# newagent.build — full content > Build an AI agent stack layer by layer. Compare 89 components — harness, model, memory, retrieval, security, runtime — and leave with a runnable setup script. This file inlines the home overview, every layer, every component, every curated stack and every guide on the site. For a shorter index with links only, see /llms.txt. # newagent.build Build an AI agent stack layer by layer. Compare 89 components — harness, model, memory, retrieval, security, runtime — and leave with a runnable setup script. ## The ten layers 1. **Harness** (What thinks) — The loop that runs your agent. Take one off the shelf, or skip it and write the loop yourself. 13 options. Required. https://newagent.build/layers/harness 2. **Model provider** (What thinks) — Where the tokens come from. A frontier API, a gateway you control, or your own GPUs. 13 options. Required. https://newagent.build/layers/model 3. **Framework** (What thinks) — The library you write orchestration in, when an off-the-shelf harness is not enough. 8 options. Optional. https://newagent.build/layers/framework 4. **Memory** (What it knows) — What the agent still knows tomorrow. 7 options. Optional. https://newagent.build/layers/memory 5. **Retrieval** (What it knows) — Your own corpus, made searchable. 6 options. Optional. https://newagent.build/layers/retrieval 6. **Web grounding** (What it knows) — Live information from outside your walls, in a shape a model can read. 5 options. Optional. https://newagent.build/layers/search 7. **Tools & integrations** (What it touches) — How the agent acts inside other people's software, and who holds the token when it does. 6 options. Optional. https://newagent.build/layers/tools 8. **Security** (What stops it) — Two halves, and you want both. Guardrails decide what the agent may say and do — injection defence, PII redaction, identity. Containment decides what holds when the guardrails do not: a kernel boundary around code it wrote, a default-deny boundary around everything it can reach. Egress control does not stop prompt injection; it removes the payout. 17 options. Required. https://newagent.build/layers/security 9. **Observability & evals** (How you run it) — Traces of what it did, and a score for whether that was any good. 6 options. Required. https://newagent.build/layers/observability 10. **Runtime & deploy** (How you run it) — Where it runs, and whether it survives a restart. 8 options. Required. https://newagent.build/layers/runtime ## Start here - Describe what you are building: `GET https://newagent.build/api/advise?q=...` - Browse everything: https://newagent.build/providers - Curated stacks: https://newagent.build/stacks - Guides: https://newagent.build/guides ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # The ten layers # Harness for AI agents The loop that runs your agent. Take one off the shelf, or skip it and write the loop yourself. 13 options tracked · 11 open source · 9 self-hostable · required ## All options | Option | What it does | Licence | Self-host | Page | | --- | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | https://newagent.build/c/claude-code | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | https://newagent.build/c/claude-agent-sdk | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | https://newagent.build/c/hermes-agent | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | https://newagent.build/c/openclaw | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | https://newagent.build/c/omnigent | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | https://newagent.build/c/prime-agent | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | https://newagent.build/c/deepseek-harness | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | https://newagent.build/c/openai-agents-sdk | | Codex CLI | OpenAI's terminal coding agent. | open source | no | https://newagent.build/c/codex-cli | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | https://newagent.build/c/opencode | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | https://newagent.build/c/goose | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | https://newagent.build/c/letta | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | https://newagent.build/c/build-your-own | ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Model provider for AI agents Where the tokens come from. A frontier API, a gateway you control, or your own GPUs. 13 options tracked · 3 open source · 3 self-hostable · required ## All options | Option | What it does | Licence | Self-host | Page | | --- | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | https://newagent.build/c/anthropic | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | https://newagent.build/c/openai | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | https://newagent.build/c/google-gemini | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | https://newagent.build/c/bedrock | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | https://newagent.build/c/vertex | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | https://newagent.build/c/deepseek | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | https://newagent.build/c/openrouter | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | https://newagent.build/c/huggingface | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | https://newagent.build/c/venice | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | https://newagent.build/c/litellm | | Groq | Open-weight models at very low latency. | proprietary | no | https://newagent.build/c/groq | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | https://newagent.build/c/ollama | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | https://newagent.build/c/vllm | ## Benchmark: SWE-bench Verified Share of real, human-validated GitHub issues resolved end to end. - Anthropic: 96% — leaderboard, SWE-bench Verified leaderboard, 2026-08-18. https://benchlm.ai/benchmarks/sweVerified - DeepSeek: 96.4% — leaderboard, SWE-bench Verified leaderboard, 2026-08-18. https://benchlm.ai/benchmarks/sweVerified > Scores attach to a specific model, not to a provider — the model measured is named in each cell. Frontier results now cluster inside a single point and different leaderboards report different figures for the same model, so treat anything under ~1 point as noise rather than a ranking. ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Framework for AI agents The library you write orchestration in, when an off-the-shelf harness is not enough. 8 options tracked · 8 open source · 8 self-hostable · optional ## All options | Option | What it does | Licence | Self-host | Page | | --- | --- | --- | --- | --- | | LangGraph | Graph-shaped orchestration with checkpointing, interrupts and human-in-the-loop. | open source | yes | https://newagent.build/c/langgraph | | Mastra | TypeScript agent framework with workflows, memory, evals and a local playground. | open source | yes | https://newagent.build/c/mastra | | Pydantic AI | Type-safe Python agents with structured output that actually validates. | open source | yes | https://newagent.build/c/pydantic-ai | | CrewAI | Role-based crews of agents with delegation between them. | open source | yes | https://newagent.build/c/crewai | | Agno | Fast, lightweight multi-agent runtime with built-in memory and knowledge. | open source | yes | https://newagent.build/c/agno | | LlamaIndex | Data-first framework — ingestion, indexing and RAG-heavy agents. | open source | yes | https://newagent.build/c/llamaindex | | DSPy | Programs, not prompts — compile and optimise the prompt against a metric. | open source | yes | https://newagent.build/c/dspy | | Vercel AI SDK | Streaming-first TypeScript SDK with a unified provider interface and UI hooks. | open source | yes | https://newagent.build/c/vercel-ai-sdk | ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Memory for AI agents What the agent still knows tomorrow. 7 options tracked · 6 open source · 6 self-hostable · optional ## All options | Option | What it does | Licence | Self-host | Page | | --- | --- | --- | --- | --- | | Mem0 | Extracts durable facts from conversations and recalls them on the next turn. | open source | yes | https://newagent.build/c/mem0 | | Zep | Temporal knowledge graph memory — facts with validity windows, not just embeddings. | open source | yes | https://newagent.build/c/zep | | Honcho | Treats memory as a reasoning problem, not a retrieval one — you ask its Dialectic API a question about the user and get a synthesised answer. | open source | yes | https://newagent.build/c/honcho | | Supermemory | Memory graph plus automatic user profiles, with connectors that sync from Drive, Gmail, Notion and GitHub. | proprietary | no | https://newagent.build/c/supermemory | | LangMem | Memory primitives that plug straight into LangGraph checkpoints. | open source | yes | https://newagent.build/c/langmem | | Cognee | Builds a graph + vector memory layer over your agent's history and documents. | open source | yes | https://newagent.build/c/cognee | | Postgres (rolled by hand) | A messages table and a summaries table. Boring, auditable, yours. | open source | yes | https://newagent.build/c/postgres-memory | ## Benchmark: LoCoMo Question-answering accuracy over long, multi-session conversations. - Mem0: 92.5% — self-reported, Mem0 research page, 2026-08. https://mem0.ai/research - Mem0: 62.47% — independent, Memori Labs evaluation, 2026-08. https://memorilabs.ai/docs/memori-cloud/benchmark/results/ - Zep: 79.09% — independent, Memori Labs evaluation, 2026-08. https://memorilabs.ai/docs/memori-cloud/benchmark/results/ - Zep: 58.44% — independent, Corrected evaluation, zep-papers #5, 2026-08. https://github.com/getzep/zep-papers/issues/5 - LangMem: 78.05% — independent, Memori Labs evaluation, 2026-08. https://memorilabs.ai/docs/memori-cloud/benchmark/results/ - Letta: 83.2% — self-reported, Reported in comparative write-up, 2026-08. https://mem0.ai/blog/benchmarked-openai-memory-vs-langmem-vs-memgpt-vs-mem0-for-long-term-memory-here-s-how-they-stacked-up > Treat every figure here as contested. Most headline LoCoMo numbers are produced by the vendor being measured, and independent re-runs have differed from vendor claims by as much as 30 points on the same benchmark name. The spread between rows is the story, not the winner. ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Retrieval for AI agents Your own corpus, made searchable. 6 options tracked · 4 open source · 4 self-hostable · optional ## All options | Option | What it does | Licence | Self-host | Page | | --- | --- | --- | --- | --- | | pgvector | Vector search inside the Postgres you already run. | open source | yes | https://newagent.build/c/pgvector | | Qdrant | Rust vector database with strong payload filtering, self-host or cloud. | open source | yes | https://newagent.build/c/qdrant | | Pinecone | Fully managed vector database. No index to operate. | proprietary | no | https://newagent.build/c/pinecone | | Chroma | Embedded vector store that runs in-process. Ideal first index. | open source | yes | https://newagent.build/c/chroma | | turbopuffer | Vector and full-text search on object storage — very cheap at rest. | proprietary | no | https://newagent.build/c/turbopuffer | | LanceDB | Embedded multimodal store backed by the Lance columnar format. | open source | yes | https://newagent.build/c/lancedb | ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Web grounding for AI agents Live information from outside your walls, in a shape a model can read. 5 options tracked · 2 open source · 2 self-hostable · optional ## All options | Option | What it does | Licence | Self-host | Page | | --- | --- | --- | --- | --- | | Exa | Embeddings-native web search built for agents, with full-content retrieval. | proprietary | no | https://newagent.build/c/exa | | Tavily | Search API that returns model-ready answers rather than ten blue links. | proprietary | no | https://newagent.build/c/tavily | | Firecrawl | Turns any site into clean markdown, including the JavaScript-heavy ones. | open source | yes | https://newagent.build/c/firecrawl | | Brave Search API | Independent web index with straightforward per-query pricing. | proprietary | no | https://newagent.build/c/brave-search | | SearXNG | Self-hosted metasearch. No third-party search vendor in the path. | open source | yes | https://newagent.build/c/searxng | ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Tools & integrations for AI agents How the agent acts inside other people's software, and who holds the token when it does. 6 options tracked · 5 open source · 4 self-hostable · optional ## All options | Option | What it does | Licence | Self-host | Page | | --- | --- | --- | --- | --- | | MCP (Model Context Protocol) | The open protocol for exposing tools and data to agents. The default answer. | open source | yes | https://newagent.build/c/mcp | | Composio | Managed auth and tool-calling for 250+ SaaS apps. | open source | no | https://newagent.build/c/composio | | Arcade | Tool-calling with per-user OAuth, so the agent acts as the user, not as root. | open source | yes | https://newagent.build/c/arcade | | Browserbase | Managed headless browsers with stealth and session replay, plus Stagehand. | proprietary | no | https://newagent.build/c/browserbase | | Playwright MCP | Drive a real local browser over MCP, using the accessibility tree. | open source | yes | https://newagent.build/c/playwright-mcp | | Nango | Open-source integration platform — you own the OAuth and the sync logic. | open source | yes | https://newagent.build/c/nango | ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Security for AI agents Two halves, and you want both. Guardrails decide what the agent may say and do — injection defence, PII redaction, identity. Containment decides what holds when the guardrails do not: a kernel boundary around code it wrote, a default-deny boundary around everything it can reach. Egress control does not stop prompt injection; it removes the payout. 17 options tracked · 13 open source · 13 self-hostable · required · pick several, they compose ## All options | Option | What it does | Licence | Self-host | Page | | --- | --- | --- | --- | --- | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | https://newagent.build/c/llama-guard | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | https://newagent.build/c/guardrails-ai | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | https://newagent.build/c/nemo-guardrails | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | https://newagent.build/c/lakera | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | https://newagent.build/c/rebuff | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | https://newagent.build/c/invariant | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | https://newagent.build/c/descope-agentic | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | https://newagent.build/c/auth0-ai | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | https://newagent.build/c/oso | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | https://newagent.build/c/presidio | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | https://newagent.build/c/cf-ai-gateway | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | https://newagent.build/c/iron-proxy | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | https://newagent.build/c/gvisor | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | https://newagent.build/c/firecracker | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | https://newagent.build/c/kata | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | https://newagent.build/c/host-only | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | https://newagent.build/c/e2b | ## Benchmark: Agent Data Injection — attack success rate Share of injection attacks that still succeed with the guardrail in place. Lower is better. - Llama Guard: 50% — peer-reviewed, Agent Data Injection Attacks, 2026-07. https://arxiv.org/pdf/2607.05120 ## Benchmark: Agent Data Injection — detection rate Share of instruction-injection attempts the guardrail flags. - Llama Guard: 34.9% — peer-reviewed, Agent Data Injection Attacks, 2026-07. https://arxiv.org/pdf/2607.05120 ## Benchmark: Recall @ 1% FPR (English) Injections caught while holding false positives to 1% — the metric that decides whether a filter is usable in production. - Llama Guard: 97.5% — self-reported, LlamaFirewall (Meta), 2026-05. https://arxiv.org/pdf/2505.03574 ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Observability & evals for AI agents Traces of what it did, and a score for whether that was any good. 6 options tracked · 4 open source · 4 self-hostable · required ## All options | Option | What it does | Licence | Self-host | Page | | --- | --- | --- | --- | --- | | Langfuse | Open-source tracing, prompt management and evals. Self-hostable in one compose file. | open source | yes | https://newagent.build/c/langfuse | | LangSmith | Tracing, datasets and evals from the LangChain team. | proprietary | no | https://newagent.build/c/langsmith | | Braintrust | Eval-first platform — scorers, datasets and a playground for prompt iteration. | proprietary | yes | https://newagent.build/c/braintrust | | Arize Phoenix | OpenTelemetry-native tracing and evals you can run locally. | open source | yes | https://newagent.build/c/phoenix | | Pydantic Logfire | OpenTelemetry observability with first-class Python and Pydantic AI support. | open source | no | https://newagent.build/c/logfire | | promptfoo | Local eval and red-team harness that runs in CI. No account needed. | open source | yes | https://newagent.build/c/promptfoo | ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Runtime & deploy for AI agents Where it runs, and whether it survives a restart. 8 options tracked · 4 open source · 4 self-hostable · required ## All options | Option | What it does | Licence | Self-host | Page | | --- | --- | --- | --- | --- | | Docker Compose | One file, one box, everything local. The honest starting point. | open source | yes | https://newagent.build/c/docker-compose | | Vercel | Streaming-friendly serverless for TypeScript agents with a web front end. | proprietary | no | https://newagent.build/c/vercel | | Cloudflare Workers | Edge runtime with Durable Objects for long-lived, stateful agent sessions. | proprietary | no | https://newagent.build/c/cloudflare-workers | | Modal | Python-native serverless with GPUs and sandboxes, scale to zero. | proprietary | no | https://newagent.build/c/modal | | Fly.io | Long-running containers close to users, without owning a Kubernetes cluster. | proprietary | no | https://newagent.build/c/fly | | Temporal | Durable execution — the run survives a crash, a deploy and a week of waiting. | open source | yes | https://newagent.build/c/temporal | | Centaur | Paradigm's self-hosted control plane: one shared team agent, an isolated k8s sandbox per Slack thread, and no raw API keys in the agent's environment. | open source | yes | https://newagent.build/c/centaur | | Kubernetes | Your existing cluster. Nothing new for the platform team to approve. | open source | yes | https://newagent.build/c/k8s | ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Every component # Claude Code Claude Code is a harness option for building AI agents. Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. ## Facts - Layer: Harness (What thinks) - Vendor: Anthropic - Licence: proprietary - Pricing: paid - Self-hostable: no - Install: `npm install -g @anthropic-ai/claude-code` - Requires: `ANTHROPIC_API_KEY` - Docs: https://claude.com/claude-code - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Worth knowing Configure via .claude/settings.json; drop reusable procedures in .claude/skills/. ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/claude-code-vs-claude-agent-sdk - https://newagent.build/compare/claude-code-vs-hermes-agent - https://newagent.build/compare/claude-code-vs-openclaw - https://newagent.build/compare/claude-code-vs-omnigent - https://newagent.build/compare/claude-code-vs-prime-agent - https://newagent.build/compare/claude-code-vs-deepseek-harness - https://newagent.build/compare/claude-code-vs-openai-agents-sdk - https://newagent.build/compare/claude-code-vs-codex-cli ## Curated stacks using it - Weekend prototype: Smallest thing that is still honest. Local, cheap, no accounts beyond one model key. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=claude-code&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Claude Agent SDK Claude Agent SDK is a harness option for building AI agents. The Claude Code loop as a library, for agents that are not about code. ## Facts - Layer: Harness (What thinks) - Vendor: Anthropic - Licence: proprietary - Pricing: usage based - Self-hostable: no - SDKs: ts, py - Install: `npm install @anthropic-ai/claude-agent-sdk` - Requires: `ANTHROPIC_API_KEY` - Docs: https://docs.claude.com/en/api/agent-sdk/overview - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Worth knowing Same harness as Claude Code, but you own the entrypoint and the tool surface. ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/claude-agent-sdk-vs-claude-code - https://newagent.build/compare/claude-agent-sdk-vs-hermes-agent - https://newagent.build/compare/claude-agent-sdk-vs-openclaw - https://newagent.build/compare/claude-agent-sdk-vs-omnigent - https://newagent.build/compare/claude-agent-sdk-vs-prime-agent - https://newagent.build/compare/claude-agent-sdk-vs-deepseek-harness - https://newagent.build/compare/claude-agent-sdk-vs-openai-agents-sdk - https://newagent.build/compare/claude-agent-sdk-vs-codex-cli ## Curated stacks using it - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=claude-agent-sdk&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Hermes Agent Hermes Agent is a harness option for building AI agents. Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. ## Facts - Layer: Harness (What thinks) - Vendor: Nous Research (github.com/NousResearch/hermes-agent) - Licence: open source - Pricing: open source - Self-hostable: yes - Install: `curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash` - Docs: https://github.com/NousResearch/hermes-agent - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing The learning loop is the point: it creates skills after complex tasks and searches its own past sessions. Runs anywhere from a $5 VPS to Modal or Vercel Sandbox. ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/hermes-agent-vs-claude-code - https://newagent.build/compare/hermes-agent-vs-claude-agent-sdk - https://newagent.build/compare/hermes-agent-vs-openclaw - https://newagent.build/compare/hermes-agent-vs-omnigent - https://newagent.build/compare/hermes-agent-vs-prime-agent - https://newagent.build/compare/hermes-agent-vs-deepseek-harness - https://newagent.build/compare/hermes-agent-vs-openai-agents-sdk - https://newagent.build/compare/hermes-agent-vs-codex-cli ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=hermes-agent&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # OpenClaw OpenClaw is a harness option for building AI agents. Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. ## Facts - Layer: Harness (What thinks) - Vendor: OpenClaw Foundation - Licence: open source - Pricing: open source - Self-hostable: yes - Install: `curl -fsSL https://openclaw.ai/install.sh | bash` - Docs: https://docs.openclaw.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing One Gateway process fans out to every channel plugin. Needs Node 22.22.3+ (26 recommended); the installer provisions it. ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/openclaw-vs-claude-code - https://newagent.build/compare/openclaw-vs-claude-agent-sdk - https://newagent.build/compare/openclaw-vs-hermes-agent - https://newagent.build/compare/openclaw-vs-omnigent - https://newagent.build/compare/openclaw-vs-prime-agent - https://newagent.build/compare/openclaw-vs-deepseek-harness - https://newagent.build/compare/openclaw-vs-openai-agents-sdk - https://newagent.build/compare/openclaw-vs-codex-cli ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=openclaw&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Omnigent Omnigent is a harness option for building AI agents. Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. ## Facts - Layer: Harness (What thinks) - Vendor: Omnigent (github.com/omnigent-ai/omnigent) - Licence: open source - Pricing: open source - Self-hostable: yes - Install: `curl -fsSL https://raw.githubusercontent.com/omnigent-ai/omnigent/main/scripts/install_oss.sh | sh` - Requires: `ANTHROPIC_API_KEY`, `OPENAI_API_KEY` - Docs: https://omnigent.ai - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing A common orchestration layer over other harnesses rather than a harness of its own — swap Claude Code/Codex/Cursor/etc. in and out, enforce policies, and run sessions in cloud sandboxes (Modal, Daytona, E2B and others). ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/omnigent-vs-claude-code - https://newagent.build/compare/omnigent-vs-claude-agent-sdk - https://newagent.build/compare/omnigent-vs-hermes-agent - https://newagent.build/compare/omnigent-vs-openclaw - https://newagent.build/compare/omnigent-vs-prime-agent - https://newagent.build/compare/omnigent-vs-deepseek-harness - https://newagent.build/compare/omnigent-vs-openai-agents-sdk - https://newagent.build/compare/omnigent-vs-codex-cli ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=omnigent&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Prime Agent Prime Agent is a harness option for building AI agents. Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. ## Facts - Layer: Harness (What thinks) - Vendor: Prime Intellect - Licence: open source - Pricing: open source - Self-hostable: yes - Install: `curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh` - Docs: https://www.primeintellect.ai/blog/prime-agent - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing Drops fixed tool schemas and context compaction: the agent writes programs over its own context, and can CRUD its own prompts, skills and sub-agents mid-run. ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/prime-agent-vs-claude-code - https://newagent.build/compare/prime-agent-vs-claude-agent-sdk - https://newagent.build/compare/prime-agent-vs-hermes-agent - https://newagent.build/compare/prime-agent-vs-openclaw - https://newagent.build/compare/prime-agent-vs-omnigent - https://newagent.build/compare/prime-agent-vs-deepseek-harness - https://newagent.build/compare/prime-agent-vs-openai-agents-sdk - https://newagent.build/compare/prime-agent-vs-codex-cli ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=prime-agent&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # DeepSeek Harness DeepSeek Harness is a harness option for building AI agents. Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. ## Facts - Layer: Harness (What thinks) - Vendor: DeepSeek AI (github.com/deepseek-ai/deepseek-harness) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: ts - Install: `npx @deepseek-ai/dsh web` - Docs: https://github.com/deepseek-ai/deepseek-harness - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing Web UI first — `dsh web` serves on 127.0.0.1:3080 rather than taking over the terminal. Built on Cordis. Still a developer preview: DeepSeek warns of compatibility-breaking changes, so pin the version. ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/deepseek-harness-vs-claude-code - https://newagent.build/compare/deepseek-harness-vs-claude-agent-sdk - https://newagent.build/compare/deepseek-harness-vs-hermes-agent - https://newagent.build/compare/deepseek-harness-vs-openclaw - https://newagent.build/compare/deepseek-harness-vs-omnigent - https://newagent.build/compare/deepseek-harness-vs-prime-agent - https://newagent.build/compare/deepseek-harness-vs-openai-agents-sdk - https://newagent.build/compare/deepseek-harness-vs-codex-cli ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=deepseek-harness&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # OpenAI Agents SDK OpenAI Agents SDK is a harness option for building AI agents. Lightweight loop with handoffs, guardrails and sessions. ## Facts - Layer: Harness (What thinks) - Vendor: OpenAI (github.com/openai/openai-agents-python) - Licence: open source - Pricing: usage based - Self-hostable: no - SDKs: py, ts - Install: `pip install openai-agents` - Requires: `OPENAI_API_KEY` - Docs: https://openai.github.io/openai-agents-python/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/openai-agents-sdk-vs-claude-code - https://newagent.build/compare/openai-agents-sdk-vs-claude-agent-sdk - https://newagent.build/compare/openai-agents-sdk-vs-hermes-agent - https://newagent.build/compare/openai-agents-sdk-vs-openclaw - https://newagent.build/compare/openai-agents-sdk-vs-omnigent - https://newagent.build/compare/openai-agents-sdk-vs-prime-agent - https://newagent.build/compare/openai-agents-sdk-vs-deepseek-harness - https://newagent.build/compare/openai-agents-sdk-vs-codex-cli ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=openai-agents-sdk&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Codex CLI Codex CLI is a harness option for building AI agents. OpenAI's terminal coding agent. ## Facts - Layer: Harness (What thinks) - Vendor: OpenAI (github.com/openai/openai-agents-python) - Licence: open source - Pricing: paid - Self-hostable: no - Install: `npm install -g @openai/codex` - Requires: `OPENAI_API_KEY` - Docs: https://developers.openai.com/codex/cli/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/codex-cli-vs-claude-code - https://newagent.build/compare/codex-cli-vs-claude-agent-sdk - https://newagent.build/compare/codex-cli-vs-hermes-agent - https://newagent.build/compare/codex-cli-vs-openclaw - https://newagent.build/compare/codex-cli-vs-omnigent - https://newagent.build/compare/codex-cli-vs-prime-agent - https://newagent.build/compare/codex-cli-vs-deepseek-harness - https://newagent.build/compare/codex-cli-vs-openai-agents-sdk ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=codex-cli&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # OpenCode OpenCode is a harness option for building AI agents. Open-source terminal coding agent, provider-agnostic. ## Facts - Layer: Harness (What thinks) - Vendor: SST (github.com/sst/opencode) - Licence: open source - Pricing: open source - Self-hostable: yes - Install: `npm install -g opencode-ai` - Docs: https://opencode.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/opencode-vs-claude-code - https://newagent.build/compare/opencode-vs-claude-agent-sdk - https://newagent.build/compare/opencode-vs-hermes-agent - https://newagent.build/compare/opencode-vs-openclaw - https://newagent.build/compare/opencode-vs-omnigent - https://newagent.build/compare/opencode-vs-prime-agent - https://newagent.build/compare/opencode-vs-deepseek-harness - https://newagent.build/compare/opencode-vs-openai-agents-sdk ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=opencode&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Goose Goose is a harness option for building AI agents. Block's open-source on-machine agent with an extension system. ## Facts - Layer: Harness (What thinks) - Vendor: Block (github.com/block/goose) - Licence: open source - Pricing: open source - Self-hostable: yes - Install: `brew install block-goose-cli` - Docs: https://block.github.io/goose/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/goose-vs-claude-code - https://newagent.build/compare/goose-vs-claude-agent-sdk - https://newagent.build/compare/goose-vs-hermes-agent - https://newagent.build/compare/goose-vs-openclaw - https://newagent.build/compare/goose-vs-omnigent - https://newagent.build/compare/goose-vs-prime-agent - https://newagent.build/compare/goose-vs-deepseek-harness - https://newagent.build/compare/goose-vs-openai-agents-sdk ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=goose&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Letta Letta is a harness option for building AI agents. Stateful agent server where memory is the primary abstraction. ## Facts - Layer: Harness (What thinks) - Vendor: Letta (github.com/letta-ai/letta) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Install: `pip install letta` - Docs: https://www.letta.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Published benchmark scores - LoCoMo: 83.2% — self-reported, Reported in comparative write-up, 2026-08. https://mem0.ai/blog/benchmarked-openai-memory-vs-langmem-vs-memgpt-vs-mem0-for-long-term-memory-here-s-how-they-stacked-up ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Roll your own loop | No harness. A while-loop, a model call and your own tool dispatch. | open source | yes | Head to head: - https://newagent.build/compare/letta-vs-claude-code - https://newagent.build/compare/letta-vs-claude-agent-sdk - https://newagent.build/compare/letta-vs-hermes-agent - https://newagent.build/compare/letta-vs-openclaw - https://newagent.build/compare/letta-vs-omnigent - https://newagent.build/compare/letta-vs-prime-agent - https://newagent.build/compare/letta-vs-deepseek-harness - https://newagent.build/compare/letta-vs-openai-agents-sdk ## Curated stacks using it - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=letta&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Roll your own loop Roll your own loop is a harness option for building AI agents. No harness. A while-loop, a model call and your own tool dispatch. ## Facts - Layer: Harness (What thinks) - Vendor: OWASP - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: ts, py - Docs: https://docs.claude.com/en/docs/agents-and-tools/tool-use/overview - Last verified: 2026-08-21 ## When to reach for it - You are choosing a harness component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing Pick a framework below, or call the model API directly and dispatch tools yourself. ## Alternatives in the harness layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Claude Code | Anthropic's coding agent — terminal, IDE and SDK, with hooks, skills and MCP built in. | proprietary | no | | Claude Agent SDK | The Claude Code loop as a library, for agents that are not about code. | proprietary | no | | Hermes Agent | Self-improving agent from Nous Research — writes its own skills from experience and carries them forward. | open source | yes | | OpenClaw | Self-hosted gateway putting an agent in Slack, Telegram, WhatsApp, Signal, Discord and iMessage at once. | open source | yes | | Omnigent | Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, OpenCode, Hermes and Pi from one session, on any device. | open source | yes | | Prime Agent | Prime Intellect's recursive-language-model harness — a persistent IPython kernel is the only tool, sub-agents are function calls. | open source | yes | | DeepSeek Harness | Everything is a plugin — the model adapter, the tool registry, the sandbox and the agent loop itself are all swappable. | open source | yes | | OpenAI Agents SDK | Lightweight loop with handoffs, guardrails and sessions. | open source | no | | Codex CLI | OpenAI's terminal coding agent. | open source | no | | OpenCode | Open-source terminal coding agent, provider-agnostic. | open source | yes | | Goose | Block's open-source on-machine agent with an extension system. | open source | yes | | Letta | Stateful agent server where memory is the primary abstraction. | open source | yes | Head to head: - https://newagent.build/compare/build-your-own-vs-claude-code - https://newagent.build/compare/build-your-own-vs-claude-agent-sdk - https://newagent.build/compare/build-your-own-vs-hermes-agent - https://newagent.build/compare/build-your-own-vs-openclaw - https://newagent.build/compare/build-your-own-vs-omnigent - https://newagent.build/compare/build-your-own-vs-prime-agent - https://newagent.build/compare/build-your-own-vs-deepseek-harness - https://newagent.build/compare/build-your-own-vs-openai-agents-sdk ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?harness=build-your-own&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Anthropic Anthropic is a model provider option for building AI agents. Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. ## Facts - Layer: Model provider (What thinks) - Vendor: Anthropic - Licence: proprietary - Pricing: usage based - Self-hostable: no - SDKs: ts, py - Install: `npm install @anthropic-ai/sdk` - Requires: `ANTHROPIC_API_KEY` - Docs: https://docs.claude.com/en/docs/about-claude/models - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Worth knowing Enable prompt caching on the system prompt and tool definitions — it is the single biggest cost lever. ## Published benchmark scores - SWE-bench Verified: 96% — leaderboard, SWE-bench Verified leaderboard, 2026-08-18. https://benchlm.ai/benchmarks/sweVerified ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Groq | Open-weight models at very low latency. | proprietary | no | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/anthropic-vs-openai - https://newagent.build/compare/anthropic-vs-google-gemini - https://newagent.build/compare/anthropic-vs-bedrock - https://newagent.build/compare/anthropic-vs-vertex - https://newagent.build/compare/anthropic-vs-deepseek - https://newagent.build/compare/anthropic-vs-openrouter - https://newagent.build/compare/anthropic-vs-huggingface - https://newagent.build/compare/anthropic-vs-venice ## Curated stacks using it - Weekend prototype: Smallest thing that is still honest. Local, cheap, no accounts beyond one model key. - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=anthropic&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # OpenAI OpenAI is a model provider option for building AI agents. GPT and o-series via the Responses API. ## Facts - Layer: Model provider (What thinks) - Vendor: OpenAI (github.com/openai/openai-agents-python) - Licence: proprietary - Pricing: usage based - Self-hostable: no - SDKs: ts, py - Install: `npm install openai` - Requires: `OPENAI_API_KEY` - Docs: https://platform.openai.com/docs/models - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Groq | Open-weight models at very low latency. | proprietary | no | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/openai-vs-anthropic - https://newagent.build/compare/openai-vs-google-gemini - https://newagent.build/compare/openai-vs-bedrock - https://newagent.build/compare/openai-vs-vertex - https://newagent.build/compare/openai-vs-deepseek - https://newagent.build/compare/openai-vs-openrouter - https://newagent.build/compare/openai-vs-huggingface - https://newagent.build/compare/openai-vs-venice ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=openai&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Google Gemini Google Gemini is a model provider option for building AI agents. Gemini models with very long context and native multimodality. ## Facts - Layer: Model provider (What thinks) - Vendor: Google - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: ts, py - Install: `npm install @google/genai` - Requires: `GOOGLE_API_KEY` - Docs: https://ai.google.dev/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Groq | Open-weight models at very low latency. | proprietary | no | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/google-gemini-vs-anthropic - https://newagent.build/compare/google-gemini-vs-openai - https://newagent.build/compare/google-gemini-vs-bedrock - https://newagent.build/compare/google-gemini-vs-vertex - https://newagent.build/compare/google-gemini-vs-deepseek - https://newagent.build/compare/google-gemini-vs-openrouter - https://newagent.build/compare/google-gemini-vs-huggingface - https://newagent.build/compare/google-gemini-vs-venice ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=google-gemini&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Amazon Bedrock Amazon Bedrock is a model provider option for building AI agents. Frontier models inside your AWS account, with IAM and VPC boundaries. ## Facts - Layer: Model provider (What thinks) - Vendor: Amazon Web Services (github.com/firecracker-microvm/firecracker) - Licence: proprietary - Pricing: usage based - Self-hostable: no - SDKs: ts, py - Requires: `AWS_REGION`, `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY` - Docs: https://aws.amazon.com/bedrock/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Worth knowing Use when procurement needs the model call to stay inside your cloud account. ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Groq | Open-weight models at very low latency. | proprietary | no | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/bedrock-vs-anthropic - https://newagent.build/compare/bedrock-vs-openai - https://newagent.build/compare/bedrock-vs-google-gemini - https://newagent.build/compare/bedrock-vs-vertex - https://newagent.build/compare/bedrock-vs-deepseek - https://newagent.build/compare/bedrock-vs-openrouter - https://newagent.build/compare/bedrock-vs-huggingface - https://newagent.build/compare/bedrock-vs-venice ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=bedrock&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Google Vertex AI Google Vertex AI is a model provider option for building AI agents. Claude and Gemini under GCP billing, IAM and regional controls. ## Facts - Layer: Model provider (What thinks) - Vendor: Google - Licence: proprietary - Pricing: usage based - Self-hostable: no - SDKs: ts, py - Requires: `GOOGLE_CLOUD_PROJECT`, `GOOGLE_APPLICATION_CREDENTIALS` - Docs: https://cloud.google.com/vertex-ai - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Groq | Open-weight models at very low latency. | proprietary | no | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/vertex-vs-anthropic - https://newagent.build/compare/vertex-vs-openai - https://newagent.build/compare/vertex-vs-google-gemini - https://newagent.build/compare/vertex-vs-bedrock - https://newagent.build/compare/vertex-vs-deepseek - https://newagent.build/compare/vertex-vs-openrouter - https://newagent.build/compare/vertex-vs-huggingface - https://newagent.build/compare/vertex-vs-venice ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=vertex&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # DeepSeek DeepSeek is a model provider option for building AI agents. Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. ## Facts - Layer: Model provider (What thinks) - Vendor: DeepSeek AI (github.com/deepseek-ai/deepseek-harness) - Licence: proprietary - Pricing: usage based - Self-hostable: no - SDKs: ts, py - Requires: `DEEPSEEK_API_KEY` - Docs: https://api-docs.deepseek.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Worth knowing OpenAI-compatible, so it is a base-URL swap. The open weights are the real hedge: if the hosted API is ever unacceptable, the same model runs on vLLM. ## Published benchmark scores - SWE-bench Verified: 96.4% — leaderboard, SWE-bench Verified leaderboard, 2026-08-18. https://benchlm.ai/benchmarks/sweVerified ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Groq | Open-weight models at very low latency. | proprietary | no | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/deepseek-vs-anthropic - https://newagent.build/compare/deepseek-vs-openai - https://newagent.build/compare/deepseek-vs-google-gemini - https://newagent.build/compare/deepseek-vs-bedrock - https://newagent.build/compare/deepseek-vs-vertex - https://newagent.build/compare/deepseek-vs-openrouter - https://newagent.build/compare/deepseek-vs-huggingface - https://newagent.build/compare/deepseek-vs-venice ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=deepseek&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # OpenRouter OpenRouter is a model provider option for building AI agents. One API key, several hundred models, automatic failover. ## Facts - Layer: Model provider (What thinks) - Vendor: OpenRouter - Licence: proprietary - Pricing: usage based - Self-hostable: no - SDKs: ts, py - Requires: `OPENROUTER_API_KEY` - Docs: https://openrouter.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Groq | Open-weight models at very low latency. | proprietary | no | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/openrouter-vs-anthropic - https://newagent.build/compare/openrouter-vs-openai - https://newagent.build/compare/openrouter-vs-google-gemini - https://newagent.build/compare/openrouter-vs-bedrock - https://newagent.build/compare/openrouter-vs-vertex - https://newagent.build/compare/openrouter-vs-deepseek - https://newagent.build/compare/openrouter-vs-huggingface - https://newagent.build/compare/openrouter-vs-venice ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=openrouter&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Hugging Face Hugging Face is a model provider option for building AI agents. One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. ## Facts - Layer: Model provider (What thinks) - Vendor: Hugging Face (github.com/huggingface/huggingface_hub) - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: ts, py - Install: `pip install huggingface_hub` - Requires: `HF_TOKEN` - Docs: https://huggingface.co/docs/inference-providers/index - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Worth knowing Two modes: let Hugging Face route and bill, or bring your own provider key and use it purely as a client. Inference Endpoints is the sibling product when you want a dedicated scale-to-zero GPU instead. ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Groq | Open-weight models at very low latency. | proprietary | no | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/huggingface-vs-anthropic - https://newagent.build/compare/huggingface-vs-openai - https://newagent.build/compare/huggingface-vs-google-gemini - https://newagent.build/compare/huggingface-vs-bedrock - https://newagent.build/compare/huggingface-vs-vertex - https://newagent.build/compare/huggingface-vs-deepseek - https://newagent.build/compare/huggingface-vs-openrouter - https://newagent.build/compare/huggingface-vs-venice ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=huggingface&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Venice AI Venice AI is a model provider option for building AI agents. Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. ## Facts - Layer: Model provider (What thinks) - Vendor: Venice AI - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: ts, py - Requires: `VENICE_API_KEY` - Docs: https://venice.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Worth knowing OpenAI-compatible, so it is a base-URL change. Privacy here means Venice does not retain prompts — it is still a third party on the wire, so it does not satisfy a genuine on-prem or data-residency requirement. ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Groq | Open-weight models at very low latency. | proprietary | no | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/venice-vs-anthropic - https://newagent.build/compare/venice-vs-openai - https://newagent.build/compare/venice-vs-google-gemini - https://newagent.build/compare/venice-vs-bedrock - https://newagent.build/compare/venice-vs-vertex - https://newagent.build/compare/venice-vs-deepseek - https://newagent.build/compare/venice-vs-openrouter - https://newagent.build/compare/venice-vs-huggingface ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=venice&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # LiteLLM LiteLLM is a model provider option for building AI agents. Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. ## Facts - Layer: Model provider (What thinks) - Vendor: BerriAI (github.com/BerriAI/litellm) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py, ts - Install: `pip install 'litellm[proxy]'` - Docs: https://www.litellm.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing Give every agent its own virtual key so you can attribute and cap spend per surface. ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | Groq | Open-weight models at very low latency. | proprietary | no | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/litellm-vs-anthropic - https://newagent.build/compare/litellm-vs-openai - https://newagent.build/compare/litellm-vs-google-gemini - https://newagent.build/compare/litellm-vs-bedrock - https://newagent.build/compare/litellm-vs-vertex - https://newagent.build/compare/litellm-vs-deepseek - https://newagent.build/compare/litellm-vs-openrouter - https://newagent.build/compare/litellm-vs-huggingface ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=litellm&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Groq Groq is a model provider option for building AI agents. Open-weight models at very low latency. ## Facts - Layer: Model provider (What thinks) - Vendor: Groq - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: ts, py - Requires: `GROQ_API_KEY` - Docs: https://groq.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/groq-vs-anthropic - https://newagent.build/compare/groq-vs-openai - https://newagent.build/compare/groq-vs-google-gemini - https://newagent.build/compare/groq-vs-bedrock - https://newagent.build/compare/groq-vs-vertex - https://newagent.build/compare/groq-vs-deepseek - https://newagent.build/compare/groq-vs-openrouter - https://newagent.build/compare/groq-vs-huggingface ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=groq&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Ollama Ollama is a model provider option for building AI agents. Open-weight models on your own machine. Nothing leaves the box. ## Facts - Layer: Model provider (What thinks) - Vendor: Ollama (github.com/ollama/ollama) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: ts, py - Install: `brew install ollama` - Docs: https://ollama.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Groq | Open-weight models at very low latency. | proprietary | no | | vLLM | High-throughput open-weight serving on your own GPUs. | open source | yes | Head to head: - https://newagent.build/compare/ollama-vs-anthropic - https://newagent.build/compare/ollama-vs-openai - https://newagent.build/compare/ollama-vs-google-gemini - https://newagent.build/compare/ollama-vs-bedrock - https://newagent.build/compare/ollama-vs-vertex - https://newagent.build/compare/ollama-vs-deepseek - https://newagent.build/compare/ollama-vs-openrouter - https://newagent.build/compare/ollama-vs-huggingface ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=ollama&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # vLLM vLLM is a model provider option for building AI agents. High-throughput open-weight serving on your own GPUs. ## Facts - Layer: Model provider (What thinks) - Vendor: vLLM (github.com/vllm-project/vllm) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py - Install: `pip install vllm` - Docs: https://docs.vllm.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a model provider component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the model provider layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Anthropic | Claude Opus, Sonnet and Haiku. Strong tool use and long-horizon agentic work. | proprietary | no | | OpenAI | GPT and o-series via the Responses API. | proprietary | no | | Google Gemini | Gemini models with very long context and native multimodality. | proprietary | no | | Amazon Bedrock | Frontier models inside your AWS account, with IAM and VPC boundaries. | proprietary | no | | Google Vertex AI | Claude and Gemini under GCP billing, IAM and regional controls. | proprietary | no | | DeepSeek | Strong reasoning and coding at a fraction of frontier pricing, with the weights published so you can move off the API later. | proprietary | no | | OpenRouter | One API key, several hundred models, automatic failover. | proprietary | no | | Hugging Face | One OpenAI-compatible endpoint routed across Groq, Together, Fireworks, Cerebras and Replicate — with the open-weight catalogue behind it. | proprietary | no | | Venice AI | Hosted open-weight inference with no prompt logging or retention, plus an anonymising proxy in front of the frontier APIs. | proprietary | no | | LiteLLM | Self-hosted proxy that speaks one API to 100+ providers, with keys and budgets. | open source | yes | | Groq | Open-weight models at very low latency. | proprietary | no | | Ollama | Open-weight models on your own machine. Nothing leaves the box. | open source | yes | Head to head: - https://newagent.build/compare/vllm-vs-anthropic - https://newagent.build/compare/vllm-vs-openai - https://newagent.build/compare/vllm-vs-google-gemini - https://newagent.build/compare/vllm-vs-bedrock - https://newagent.build/compare/vllm-vs-vertex - https://newagent.build/compare/vllm-vs-deepseek - https://newagent.build/compare/vllm-vs-openrouter - https://newagent.build/compare/vllm-vs-huggingface ## Curated stacks using it - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?model=vllm&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # LangGraph LangGraph is a framework option for building AI agents. Graph-shaped orchestration with checkpointing, interrupts and human-in-the-loop. ## Facts - Layer: Framework (What thinks) - Vendor: LangChain (github.com/langchain-ai/langgraph) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py, ts - Install: `pip install langgraph` - Docs: https://langchain-ai.github.io/langgraph/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a framework component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the framework layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Mastra | TypeScript agent framework with workflows, memory, evals and a local playground. | open source | yes | | Pydantic AI | Type-safe Python agents with structured output that actually validates. | open source | yes | | CrewAI | Role-based crews of agents with delegation between them. | open source | yes | | Agno | Fast, lightweight multi-agent runtime with built-in memory and knowledge. | open source | yes | | LlamaIndex | Data-first framework — ingestion, indexing and RAG-heavy agents. | open source | yes | | DSPy | Programs, not prompts — compile and optimise the prompt against a metric. | open source | yes | | Vercel AI SDK | Streaming-first TypeScript SDK with a unified provider interface and UI hooks. | open source | yes | Head to head: - https://newagent.build/compare/langgraph-vs-mastra - https://newagent.build/compare/langgraph-vs-pydantic-ai - https://newagent.build/compare/langgraph-vs-crewai - https://newagent.build/compare/langgraph-vs-agno - https://newagent.build/compare/langgraph-vs-llamaindex - https://newagent.build/compare/langgraph-vs-dspy - https://newagent.build/compare/langgraph-vs-vercel-ai-sdk ## Curated stacks using it - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?framework=langgraph&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Mastra Mastra is a framework option for building AI agents. TypeScript agent framework with workflows, memory, evals and a local playground. ## Facts - Layer: Framework (What thinks) - Vendor: Mastra (github.com/mastra-ai/mastra) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: ts - Install: `npm install @mastra/core` - Docs: https://mastra.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a framework component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the framework layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | LangGraph | Graph-shaped orchestration with checkpointing, interrupts and human-in-the-loop. | open source | yes | | Pydantic AI | Type-safe Python agents with structured output that actually validates. | open source | yes | | CrewAI | Role-based crews of agents with delegation between them. | open source | yes | | Agno | Fast, lightweight multi-agent runtime with built-in memory and knowledge. | open source | yes | | LlamaIndex | Data-first framework — ingestion, indexing and RAG-heavy agents. | open source | yes | | DSPy | Programs, not prompts — compile and optimise the prompt against a metric. | open source | yes | | Vercel AI SDK | Streaming-first TypeScript SDK with a unified provider interface and UI hooks. | open source | yes | Head to head: - https://newagent.build/compare/mastra-vs-langgraph - https://newagent.build/compare/mastra-vs-pydantic-ai - https://newagent.build/compare/mastra-vs-crewai - https://newagent.build/compare/mastra-vs-agno - https://newagent.build/compare/mastra-vs-llamaindex - https://newagent.build/compare/mastra-vs-dspy - https://newagent.build/compare/mastra-vs-vercel-ai-sdk ## Curated stacks using it - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?framework=mastra&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Pydantic AI Pydantic AI is a framework option for building AI agents. Type-safe Python agents with structured output that actually validates. ## Facts - Layer: Framework (What thinks) - Vendor: Pydantic (github.com/pydantic/pydantic-ai) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py - Install: `pip install pydantic-ai` - Docs: https://ai.pydantic.dev/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a framework component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the framework layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | LangGraph | Graph-shaped orchestration with checkpointing, interrupts and human-in-the-loop. | open source | yes | | Mastra | TypeScript agent framework with workflows, memory, evals and a local playground. | open source | yes | | CrewAI | Role-based crews of agents with delegation between them. | open source | yes | | Agno | Fast, lightweight multi-agent runtime with built-in memory and knowledge. | open source | yes | | LlamaIndex | Data-first framework — ingestion, indexing and RAG-heavy agents. | open source | yes | | DSPy | Programs, not prompts — compile and optimise the prompt against a metric. | open source | yes | | Vercel AI SDK | Streaming-first TypeScript SDK with a unified provider interface and UI hooks. | open source | yes | Head to head: - https://newagent.build/compare/pydantic-ai-vs-langgraph - https://newagent.build/compare/pydantic-ai-vs-mastra - https://newagent.build/compare/pydantic-ai-vs-crewai - https://newagent.build/compare/pydantic-ai-vs-agno - https://newagent.build/compare/pydantic-ai-vs-llamaindex - https://newagent.build/compare/pydantic-ai-vs-dspy - https://newagent.build/compare/pydantic-ai-vs-vercel-ai-sdk ## Curated stacks using it - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?framework=pydantic-ai&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # CrewAI CrewAI is a framework option for building AI agents. Role-based crews of agents with delegation between them. ## Facts - Layer: Framework (What thinks) - Vendor: CrewAI (github.com/crewAIInc/crewAI) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py - Install: `pip install crewai` - Docs: https://www.crewai.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a framework component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Alternatives in the framework layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | LangGraph | Graph-shaped orchestration with checkpointing, interrupts and human-in-the-loop. | open source | yes | | Mastra | TypeScript agent framework with workflows, memory, evals and a local playground. | open source | yes | | Pydantic AI | Type-safe Python agents with structured output that actually validates. | open source | yes | | Agno | Fast, lightweight multi-agent runtime with built-in memory and knowledge. | open source | yes | | LlamaIndex | Data-first framework — ingestion, indexing and RAG-heavy agents. | open source | yes | | DSPy | Programs, not prompts — compile and optimise the prompt against a metric. | open source | yes | | Vercel AI SDK | Streaming-first TypeScript SDK with a unified provider interface and UI hooks. | open source | yes | Head to head: - https://newagent.build/compare/crewai-vs-langgraph - https://newagent.build/compare/crewai-vs-mastra - https://newagent.build/compare/crewai-vs-pydantic-ai - https://newagent.build/compare/crewai-vs-agno - https://newagent.build/compare/crewai-vs-llamaindex - https://newagent.build/compare/crewai-vs-dspy - https://newagent.build/compare/crewai-vs-vercel-ai-sdk ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?framework=crewai&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Agno Agno is a framework option for building AI agents. Fast, lightweight multi-agent runtime with built-in memory and knowledge. ## Facts - Layer: Framework (What thinks) - Vendor: Agno (github.com/agno-agi/agno) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py - Install: `pip install agno` - Docs: https://www.agno.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a framework component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the framework layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | LangGraph | Graph-shaped orchestration with checkpointing, interrupts and human-in-the-loop. | open source | yes | | Mastra | TypeScript agent framework with workflows, memory, evals and a local playground. | open source | yes | | Pydantic AI | Type-safe Python agents with structured output that actually validates. | open source | yes | | CrewAI | Role-based crews of agents with delegation between them. | open source | yes | | LlamaIndex | Data-first framework — ingestion, indexing and RAG-heavy agents. | open source | yes | | DSPy | Programs, not prompts — compile and optimise the prompt against a metric. | open source | yes | | Vercel AI SDK | Streaming-first TypeScript SDK with a unified provider interface and UI hooks. | open source | yes | Head to head: - https://newagent.build/compare/agno-vs-langgraph - https://newagent.build/compare/agno-vs-mastra - https://newagent.build/compare/agno-vs-pydantic-ai - https://newagent.build/compare/agno-vs-crewai - https://newagent.build/compare/agno-vs-llamaindex - https://newagent.build/compare/agno-vs-dspy - https://newagent.build/compare/agno-vs-vercel-ai-sdk ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?framework=agno&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # LlamaIndex LlamaIndex is a framework option for building AI agents. Data-first framework — ingestion, indexing and RAG-heavy agents. ## Facts - Layer: Framework (What thinks) - Vendor: LlamaIndex (github.com/run-llama/llama_index) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Install: `pip install llama-index` - Docs: https://www.llamaindex.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a framework component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Alternatives in the framework layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | LangGraph | Graph-shaped orchestration with checkpointing, interrupts and human-in-the-loop. | open source | yes | | Mastra | TypeScript agent framework with workflows, memory, evals and a local playground. | open source | yes | | Pydantic AI | Type-safe Python agents with structured output that actually validates. | open source | yes | | CrewAI | Role-based crews of agents with delegation between them. | open source | yes | | Agno | Fast, lightweight multi-agent runtime with built-in memory and knowledge. | open source | yes | | DSPy | Programs, not prompts — compile and optimise the prompt against a metric. | open source | yes | | Vercel AI SDK | Streaming-first TypeScript SDK with a unified provider interface and UI hooks. | open source | yes | Head to head: - https://newagent.build/compare/llamaindex-vs-langgraph - https://newagent.build/compare/llamaindex-vs-mastra - https://newagent.build/compare/llamaindex-vs-pydantic-ai - https://newagent.build/compare/llamaindex-vs-crewai - https://newagent.build/compare/llamaindex-vs-agno - https://newagent.build/compare/llamaindex-vs-dspy - https://newagent.build/compare/llamaindex-vs-vercel-ai-sdk ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?framework=llamaindex&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # DSPy DSPy is a framework option for building AI agents. Programs, not prompts — compile and optimise the prompt against a metric. ## Facts - Layer: Framework (What thinks) - Vendor: Stanford NLP (github.com/stanfordnlp/dspy) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py - Install: `pip install dspy` - Docs: https://dspy.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a framework component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the framework layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | LangGraph | Graph-shaped orchestration with checkpointing, interrupts and human-in-the-loop. | open source | yes | | Mastra | TypeScript agent framework with workflows, memory, evals and a local playground. | open source | yes | | Pydantic AI | Type-safe Python agents with structured output that actually validates. | open source | yes | | CrewAI | Role-based crews of agents with delegation between them. | open source | yes | | Agno | Fast, lightweight multi-agent runtime with built-in memory and knowledge. | open source | yes | | LlamaIndex | Data-first framework — ingestion, indexing and RAG-heavy agents. | open source | yes | | Vercel AI SDK | Streaming-first TypeScript SDK with a unified provider interface and UI hooks. | open source | yes | Head to head: - https://newagent.build/compare/dspy-vs-langgraph - https://newagent.build/compare/dspy-vs-mastra - https://newagent.build/compare/dspy-vs-pydantic-ai - https://newagent.build/compare/dspy-vs-crewai - https://newagent.build/compare/dspy-vs-agno - https://newagent.build/compare/dspy-vs-llamaindex - https://newagent.build/compare/dspy-vs-vercel-ai-sdk ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?framework=dspy&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Vercel AI SDK Vercel AI SDK is a framework option for building AI agents. Streaming-first TypeScript SDK with a unified provider interface and UI hooks. ## Facts - Layer: Framework (What thinks) - Vendor: Vercel (github.com/vercel/ai) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: ts - Install: `npm install ai` - Docs: https://ai-sdk.dev/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a framework component — what thinks. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the framework layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | LangGraph | Graph-shaped orchestration with checkpointing, interrupts and human-in-the-loop. | open source | yes | | Mastra | TypeScript agent framework with workflows, memory, evals and a local playground. | open source | yes | | Pydantic AI | Type-safe Python agents with structured output that actually validates. | open source | yes | | CrewAI | Role-based crews of agents with delegation between them. | open source | yes | | Agno | Fast, lightweight multi-agent runtime with built-in memory and knowledge. | open source | yes | | LlamaIndex | Data-first framework — ingestion, indexing and RAG-heavy agents. | open source | yes | | DSPy | Programs, not prompts — compile and optimise the prompt against a metric. | open source | yes | Head to head: - https://newagent.build/compare/vercel-ai-sdk-vs-langgraph - https://newagent.build/compare/vercel-ai-sdk-vs-mastra - https://newagent.build/compare/vercel-ai-sdk-vs-pydantic-ai - https://newagent.build/compare/vercel-ai-sdk-vs-crewai - https://newagent.build/compare/vercel-ai-sdk-vs-agno - https://newagent.build/compare/vercel-ai-sdk-vs-llamaindex - https://newagent.build/compare/vercel-ai-sdk-vs-dspy ## Curated stacks using it - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?framework=vercel-ai-sdk&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Mem0 Mem0 is a memory option for building AI agents. Extracts durable facts from conversations and recalls them on the next turn. ## Facts - Layer: Memory (What it knows) - Vendor: Mem0 (github.com/mem0ai/mem0) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Install: `pip install mem0ai` - Requires: `MEM0_API_KEY` - Docs: https://mem0.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a memory component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Published benchmark scores - LoCoMo: 92.5% — self-reported, Mem0 research page, 2026-08. https://mem0.ai/research - LoCoMo: 62.47% — independent, Memori Labs evaluation, 2026-08. https://memorilabs.ai/docs/memori-cloud/benchmark/results/ ## Alternatives in the memory layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Zep | Temporal knowledge graph memory — facts with validity windows, not just embeddings. | open source | yes | | Honcho | Treats memory as a reasoning problem, not a retrieval one — you ask its Dialectic API a question about the user and get a synthesised answer. | open source | yes | | Supermemory | Memory graph plus automatic user profiles, with connectors that sync from Drive, Gmail, Notion and GitHub. | proprietary | no | | LangMem | Memory primitives that plug straight into LangGraph checkpoints. | open source | yes | | Cognee | Builds a graph + vector memory layer over your agent's history and documents. | open source | yes | | Postgres (rolled by hand) | A messages table and a summaries table. Boring, auditable, yours. | open source | yes | Head to head: - https://newagent.build/compare/mem0-vs-zep - https://newagent.build/compare/mem0-vs-honcho - https://newagent.build/compare/mem0-vs-supermemory - https://newagent.build/compare/mem0-vs-langmem - https://newagent.build/compare/mem0-vs-cognee - https://newagent.build/compare/mem0-vs-postgres-memory ## Curated stacks using it - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?memory=mem0&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Zep Zep is a memory option for building AI agents. Temporal knowledge graph memory — facts with validity windows, not just embeddings. ## Facts - Layer: Memory (What it knows) - Vendor: Zep (github.com/getzep/zep) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Requires: `ZEP_API_KEY` - Docs: https://www.getzep.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a memory component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Published benchmark scores - LoCoMo: 79.09% — independent, Memori Labs evaluation, 2026-08. https://memorilabs.ai/docs/memori-cloud/benchmark/results/ - LoCoMo: 58.44% — independent, Corrected evaluation, zep-papers #5, 2026-08. https://github.com/getzep/zep-papers/issues/5 ## Alternatives in the memory layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Mem0 | Extracts durable facts from conversations and recalls them on the next turn. | open source | yes | | Honcho | Treats memory as a reasoning problem, not a retrieval one — you ask its Dialectic API a question about the user and get a synthesised answer. | open source | yes | | Supermemory | Memory graph plus automatic user profiles, with connectors that sync from Drive, Gmail, Notion and GitHub. | proprietary | no | | LangMem | Memory primitives that plug straight into LangGraph checkpoints. | open source | yes | | Cognee | Builds a graph + vector memory layer over your agent's history and documents. | open source | yes | | Postgres (rolled by hand) | A messages table and a summaries table. Boring, auditable, yours. | open source | yes | Head to head: - https://newagent.build/compare/zep-vs-mem0 - https://newagent.build/compare/zep-vs-honcho - https://newagent.build/compare/zep-vs-supermemory - https://newagent.build/compare/zep-vs-langmem - https://newagent.build/compare/zep-vs-cognee - https://newagent.build/compare/zep-vs-postgres-memory ## Curated stacks using it - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?memory=zep&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Honcho Honcho is a memory option for building AI agents. Treats memory as a reasoning problem, not a retrieval one — you ask its Dialectic API a question about the user and get a synthesised answer. ## Facts - Layer: Memory (What it knows) - Vendor: Plastic Labs (github.com/plastic-labs/honcho) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Docs: https://github.com/plastic-labs/honcho - Last verified: 2026-08-21 ## When to reach for it - You are choosing a memory component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Worth knowing A fine-tuned model extracts preferences, beliefs and contradictions in the background rather than embedding transcripts. Managed at api.honcho.dev, or self-host the FastAPI server. Hermes Agent uses it for user modelling. ## Alternatives in the memory layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Mem0 | Extracts durable facts from conversations and recalls them on the next turn. | open source | yes | | Zep | Temporal knowledge graph memory — facts with validity windows, not just embeddings. | open source | yes | | Supermemory | Memory graph plus automatic user profiles, with connectors that sync from Drive, Gmail, Notion and GitHub. | proprietary | no | | LangMem | Memory primitives that plug straight into LangGraph checkpoints. | open source | yes | | Cognee | Builds a graph + vector memory layer over your agent's history and documents. | open source | yes | | Postgres (rolled by hand) | A messages table and a summaries table. Boring, auditable, yours. | open source | yes | Head to head: - https://newagent.build/compare/honcho-vs-mem0 - https://newagent.build/compare/honcho-vs-zep - https://newagent.build/compare/honcho-vs-supermemory - https://newagent.build/compare/honcho-vs-langmem - https://newagent.build/compare/honcho-vs-cognee - https://newagent.build/compare/honcho-vs-postgres-memory ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?memory=honcho&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Supermemory Supermemory is a memory option for building AI agents. Memory graph plus automatic user profiles, with connectors that sync from Drive, Gmail, Notion and GitHub. ## Facts - Layer: Memory (What it knows) - Vendor: Supermemory - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: ts, py - Requires: `SUPERMEMORY_API_KEY` - Docs: https://supermemory.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a memory component — what it knows. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Worth knowing Closest thing here to memory and RAG in one product: the graph tracks contradictions and recency, and the connectors mean you are not writing an ingestion pipeline first. ## Alternatives in the memory layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Mem0 | Extracts durable facts from conversations and recalls them on the next turn. | open source | yes | | Zep | Temporal knowledge graph memory — facts with validity windows, not just embeddings. | open source | yes | | Honcho | Treats memory as a reasoning problem, not a retrieval one — you ask its Dialectic API a question about the user and get a synthesised answer. | open source | yes | | LangMem | Memory primitives that plug straight into LangGraph checkpoints. | open source | yes | | Cognee | Builds a graph + vector memory layer over your agent's history and documents. | open source | yes | | Postgres (rolled by hand) | A messages table and a summaries table. Boring, auditable, yours. | open source | yes | Head to head: - https://newagent.build/compare/supermemory-vs-mem0 - https://newagent.build/compare/supermemory-vs-zep - https://newagent.build/compare/supermemory-vs-honcho - https://newagent.build/compare/supermemory-vs-langmem - https://newagent.build/compare/supermemory-vs-cognee - https://newagent.build/compare/supermemory-vs-postgres-memory ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?memory=supermemory&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # LangMem LangMem is a memory option for building AI agents. Memory primitives that plug straight into LangGraph checkpoints. ## Facts - Layer: Memory (What it knows) - Vendor: LangChain (github.com/langchain-ai/langgraph) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py - Install: `pip install langmem` - Docs: https://langchain-ai.github.io/langmem/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a memory component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Published benchmark scores - LoCoMo: 78.05% — independent, Memori Labs evaluation, 2026-08. https://memorilabs.ai/docs/memori-cloud/benchmark/results/ ## Alternatives in the memory layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Mem0 | Extracts durable facts from conversations and recalls them on the next turn. | open source | yes | | Zep | Temporal knowledge graph memory — facts with validity windows, not just embeddings. | open source | yes | | Honcho | Treats memory as a reasoning problem, not a retrieval one — you ask its Dialectic API a question about the user and get a synthesised answer. | open source | yes | | Supermemory | Memory graph plus automatic user profiles, with connectors that sync from Drive, Gmail, Notion and GitHub. | proprietary | no | | Cognee | Builds a graph + vector memory layer over your agent's history and documents. | open source | yes | | Postgres (rolled by hand) | A messages table and a summaries table. Boring, auditable, yours. | open source | yes | Head to head: - https://newagent.build/compare/langmem-vs-mem0 - https://newagent.build/compare/langmem-vs-zep - https://newagent.build/compare/langmem-vs-honcho - https://newagent.build/compare/langmem-vs-supermemory - https://newagent.build/compare/langmem-vs-cognee - https://newagent.build/compare/langmem-vs-postgres-memory ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?memory=langmem&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Cognee Cognee is a memory option for building AI agents. Builds a graph + vector memory layer over your agent's history and documents. ## Facts - Layer: Memory (What it knows) - Vendor: Cognee (github.com/topoteretes/cognee) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py - Install: `pip install cognee` - Docs: https://www.cognee.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a memory component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the memory layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Mem0 | Extracts durable facts from conversations and recalls them on the next turn. | open source | yes | | Zep | Temporal knowledge graph memory — facts with validity windows, not just embeddings. | open source | yes | | Honcho | Treats memory as a reasoning problem, not a retrieval one — you ask its Dialectic API a question about the user and get a synthesised answer. | open source | yes | | Supermemory | Memory graph plus automatic user profiles, with connectors that sync from Drive, Gmail, Notion and GitHub. | proprietary | no | | LangMem | Memory primitives that plug straight into LangGraph checkpoints. | open source | yes | | Postgres (rolled by hand) | A messages table and a summaries table. Boring, auditable, yours. | open source | yes | Head to head: - https://newagent.build/compare/cognee-vs-mem0 - https://newagent.build/compare/cognee-vs-zep - https://newagent.build/compare/cognee-vs-honcho - https://newagent.build/compare/cognee-vs-supermemory - https://newagent.build/compare/cognee-vs-langmem - https://newagent.build/compare/cognee-vs-postgres-memory ## Curated stacks using it - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?memory=cognee&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Postgres (rolled by hand) Postgres (rolled by hand) is a memory option for building AI agents. A messages table and a summaries table. Boring, auditable, yours. ## Facts - Layer: Memory (What it knows) - Vendor: PostgreSQL - Licence: open source - Pricing: open source - Self-hostable: yes - Docs: https://www.postgresql.org/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a memory component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing Start here. Move to a memory product when you can name the query it makes fast. ## Alternatives in the memory layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Mem0 | Extracts durable facts from conversations and recalls them on the next turn. | open source | yes | | Zep | Temporal knowledge graph memory — facts with validity windows, not just embeddings. | open source | yes | | Honcho | Treats memory as a reasoning problem, not a retrieval one — you ask its Dialectic API a question about the user and get a synthesised answer. | open source | yes | | Supermemory | Memory graph plus automatic user profiles, with connectors that sync from Drive, Gmail, Notion and GitHub. | proprietary | no | | LangMem | Memory primitives that plug straight into LangGraph checkpoints. | open source | yes | | Cognee | Builds a graph + vector memory layer over your agent's history and documents. | open source | yes | Head to head: - https://newagent.build/compare/postgres-memory-vs-mem0 - https://newagent.build/compare/postgres-memory-vs-zep - https://newagent.build/compare/postgres-memory-vs-honcho - https://newagent.build/compare/postgres-memory-vs-supermemory - https://newagent.build/compare/postgres-memory-vs-langmem - https://newagent.build/compare/postgres-memory-vs-cognee ## Curated stacks using it - Weekend prototype: Smallest thing that is still honest. Local, cheap, no accounts beyond one model key. - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?memory=postgres-memory&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # pgvector pgvector is a retrieval option for building AI agents. Vector search inside the Postgres you already run. ## Facts - Layer: Retrieval (What it knows) - Vendor: pgvector (github.com/pgvector/pgvector) - Licence: open source - Pricing: open source - Self-hostable: yes - Docs: https://github.com/pgvector/pgvector - Last verified: 2026-08-21 ## When to reach for it - You are choosing a retrieval component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing Correct default under roughly ten million vectors. Reach for a dedicated store when it stops being. ## Alternatives in the retrieval layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Qdrant | Rust vector database with strong payload filtering, self-host or cloud. | open source | yes | | Pinecone | Fully managed vector database. No index to operate. | proprietary | no | | Chroma | Embedded vector store that runs in-process. Ideal first index. | open source | yes | | turbopuffer | Vector and full-text search on object storage — very cheap at rest. | proprietary | no | | LanceDB | Embedded multimodal store backed by the Lance columnar format. | open source | yes | Head to head: - https://newagent.build/compare/pgvector-vs-qdrant - https://newagent.build/compare/pgvector-vs-pinecone - https://newagent.build/compare/pgvector-vs-chroma - https://newagent.build/compare/pgvector-vs-turbopuffer - https://newagent.build/compare/pgvector-vs-lancedb ## Curated stacks using it - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?retrieval=pgvector&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Qdrant Qdrant is a retrieval option for building AI agents. Rust vector database with strong payload filtering, self-host or cloud. ## Facts - Layer: Retrieval (What it knows) - Vendor: Qdrant (github.com/qdrant/qdrant) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Docs: https://qdrant.tech/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a retrieval component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Alternatives in the retrieval layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | pgvector | Vector search inside the Postgres you already run. | open source | yes | | Pinecone | Fully managed vector database. No index to operate. | proprietary | no | | Chroma | Embedded vector store that runs in-process. Ideal first index. | open source | yes | | turbopuffer | Vector and full-text search on object storage — very cheap at rest. | proprietary | no | | LanceDB | Embedded multimodal store backed by the Lance columnar format. | open source | yes | Head to head: - https://newagent.build/compare/qdrant-vs-pgvector - https://newagent.build/compare/qdrant-vs-pinecone - https://newagent.build/compare/qdrant-vs-chroma - https://newagent.build/compare/qdrant-vs-turbopuffer - https://newagent.build/compare/qdrant-vs-lancedb ## Curated stacks using it - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?retrieval=qdrant&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Pinecone Pinecone is a retrieval option for building AI agents. Fully managed vector database. No index to operate. ## Facts - Layer: Retrieval (What it knows) - Vendor: Pinecone - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: py, ts - Requires: `PINECONE_API_KEY` - Docs: https://www.pinecone.io/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a retrieval component — what it knows. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the retrieval layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | pgvector | Vector search inside the Postgres you already run. | open source | yes | | Qdrant | Rust vector database with strong payload filtering, self-host or cloud. | open source | yes | | Chroma | Embedded vector store that runs in-process. Ideal first index. | open source | yes | | turbopuffer | Vector and full-text search on object storage — very cheap at rest. | proprietary | no | | LanceDB | Embedded multimodal store backed by the Lance columnar format. | open source | yes | Head to head: - https://newagent.build/compare/pinecone-vs-pgvector - https://newagent.build/compare/pinecone-vs-qdrant - https://newagent.build/compare/pinecone-vs-chroma - https://newagent.build/compare/pinecone-vs-turbopuffer - https://newagent.build/compare/pinecone-vs-lancedb ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?retrieval=pinecone&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Chroma Chroma is a retrieval option for building AI agents. Embedded vector store that runs in-process. Ideal first index. ## Facts - Layer: Retrieval (What it knows) - Vendor: Chroma (github.com/chroma-core/chroma) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py, ts - Install: `pip install chromadb` - Docs: https://www.trychroma.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a retrieval component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the retrieval layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | pgvector | Vector search inside the Postgres you already run. | open source | yes | | Qdrant | Rust vector database with strong payload filtering, self-host or cloud. | open source | yes | | Pinecone | Fully managed vector database. No index to operate. | proprietary | no | | turbopuffer | Vector and full-text search on object storage — very cheap at rest. | proprietary | no | | LanceDB | Embedded multimodal store backed by the Lance columnar format. | open source | yes | Head to head: - https://newagent.build/compare/chroma-vs-pgvector - https://newagent.build/compare/chroma-vs-qdrant - https://newagent.build/compare/chroma-vs-pinecone - https://newagent.build/compare/chroma-vs-turbopuffer - https://newagent.build/compare/chroma-vs-lancedb ## Curated stacks using it - Weekend prototype: Smallest thing that is still honest. Local, cheap, no accounts beyond one model key. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?retrieval=chroma&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # turbopuffer turbopuffer is a retrieval option for building AI agents. Vector and full-text search on object storage — very cheap at rest. ## Facts - Layer: Retrieval (What it knows) - Vendor: turbopuffer - Licence: proprietary - Pricing: usage based - Self-hostable: no - SDKs: py, ts - Requires: `TURBOPUFFER_API_KEY` - Docs: https://turbopuffer.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a retrieval component — what it knows. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the retrieval layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | pgvector | Vector search inside the Postgres you already run. | open source | yes | | Qdrant | Rust vector database with strong payload filtering, self-host or cloud. | open source | yes | | Pinecone | Fully managed vector database. No index to operate. | proprietary | no | | Chroma | Embedded vector store that runs in-process. Ideal first index. | open source | yes | | LanceDB | Embedded multimodal store backed by the Lance columnar format. | open source | yes | Head to head: - https://newagent.build/compare/turbopuffer-vs-pgvector - https://newagent.build/compare/turbopuffer-vs-qdrant - https://newagent.build/compare/turbopuffer-vs-pinecone - https://newagent.build/compare/turbopuffer-vs-chroma - https://newagent.build/compare/turbopuffer-vs-lancedb ## Curated stacks using it - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?retrieval=turbopuffer&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # LanceDB LanceDB is a retrieval option for building AI agents. Embedded multimodal store backed by the Lance columnar format. ## Facts - Layer: Retrieval (What it knows) - Vendor: LanceDB (github.com/lancedb/lancedb) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py, ts - Install: `pip install lancedb` - Docs: https://lancedb.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a retrieval component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the retrieval layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | pgvector | Vector search inside the Postgres you already run. | open source | yes | | Qdrant | Rust vector database with strong payload filtering, self-host or cloud. | open source | yes | | Pinecone | Fully managed vector database. No index to operate. | proprietary | no | | Chroma | Embedded vector store that runs in-process. Ideal first index. | open source | yes | | turbopuffer | Vector and full-text search on object storage — very cheap at rest. | proprietary | no | Head to head: - https://newagent.build/compare/lancedb-vs-pgvector - https://newagent.build/compare/lancedb-vs-qdrant - https://newagent.build/compare/lancedb-vs-pinecone - https://newagent.build/compare/lancedb-vs-chroma - https://newagent.build/compare/lancedb-vs-turbopuffer ## Curated stacks using it - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?retrieval=lancedb&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Exa Exa is a web grounding option for building AI agents. Embeddings-native web search built for agents, with full-content retrieval. ## Facts - Layer: Web grounding (What it knows) - Vendor: Exa - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: py, ts - Install: `npm install exa-js` - Requires: `EXA_API_KEY` - Docs: https://exa.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a web grounding component — what it knows. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the web grounding layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Tavily | Search API that returns model-ready answers rather than ten blue links. | proprietary | no | | Firecrawl | Turns any site into clean markdown, including the JavaScript-heavy ones. | open source | yes | | Brave Search API | Independent web index with straightforward per-query pricing. | proprietary | no | | SearXNG | Self-hosted metasearch. No third-party search vendor in the path. | open source | yes | Head to head: - https://newagent.build/compare/exa-vs-tavily - https://newagent.build/compare/exa-vs-firecrawl - https://newagent.build/compare/exa-vs-brave-search - https://newagent.build/compare/exa-vs-searxng ## Curated stacks using it - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?search=exa&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Tavily Tavily is a web grounding option for building AI agents. Search API that returns model-ready answers rather than ten blue links. ## Facts - Layer: Web grounding (What it knows) - Vendor: Tavily - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: py, ts - Requires: `TAVILY_API_KEY` - Docs: https://tavily.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a web grounding component — what it knows. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the web grounding layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Exa | Embeddings-native web search built for agents, with full-content retrieval. | proprietary | no | | Firecrawl | Turns any site into clean markdown, including the JavaScript-heavy ones. | open source | yes | | Brave Search API | Independent web index with straightforward per-query pricing. | proprietary | no | | SearXNG | Self-hosted metasearch. No third-party search vendor in the path. | open source | yes | Head to head: - https://newagent.build/compare/tavily-vs-exa - https://newagent.build/compare/tavily-vs-firecrawl - https://newagent.build/compare/tavily-vs-brave-search - https://newagent.build/compare/tavily-vs-searxng ## Curated stacks using it - Weekend prototype: Smallest thing that is still honest. Local, cheap, no accounts beyond one model key. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?search=tavily&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Firecrawl Firecrawl is a web grounding option for building AI agents. Turns any site into clean markdown, including the JavaScript-heavy ones. ## Facts - Layer: Web grounding (What it knows) - Vendor: Firecrawl (github.com/mendableai/firecrawl) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Requires: `FIRECRAWL_API_KEY` - Docs: https://www.firecrawl.dev/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a web grounding component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Alternatives in the web grounding layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Exa | Embeddings-native web search built for agents, with full-content retrieval. | proprietary | no | | Tavily | Search API that returns model-ready answers rather than ten blue links. | proprietary | no | | Brave Search API | Independent web index with straightforward per-query pricing. | proprietary | no | | SearXNG | Self-hosted metasearch. No third-party search vendor in the path. | open source | yes | Head to head: - https://newagent.build/compare/firecrawl-vs-exa - https://newagent.build/compare/firecrawl-vs-tavily - https://newagent.build/compare/firecrawl-vs-brave-search - https://newagent.build/compare/firecrawl-vs-searxng ## Curated stacks using it - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?search=firecrawl&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Brave Search API Brave Search API is a web grounding option for building AI agents. Independent web index with straightforward per-query pricing. ## Facts - Layer: Web grounding (What it knows) - Vendor: Brave - Licence: proprietary - Pricing: freemium - Self-hostable: no - Requires: `BRAVE_API_KEY` - Docs: https://brave.com/search/api/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a web grounding component — what it knows. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the web grounding layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Exa | Embeddings-native web search built for agents, with full-content retrieval. | proprietary | no | | Tavily | Search API that returns model-ready answers rather than ten blue links. | proprietary | no | | Firecrawl | Turns any site into clean markdown, including the JavaScript-heavy ones. | open source | yes | | SearXNG | Self-hosted metasearch. No third-party search vendor in the path. | open source | yes | Head to head: - https://newagent.build/compare/brave-search-vs-exa - https://newagent.build/compare/brave-search-vs-tavily - https://newagent.build/compare/brave-search-vs-firecrawl - https://newagent.build/compare/brave-search-vs-searxng ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?search=brave-search&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # SearXNG SearXNG is a web grounding option for building AI agents. Self-hosted metasearch. No third-party search vendor in the path. ## Facts - Layer: Web grounding (What it knows) - Vendor: SearXNG (github.com/searxng/searxng) - Licence: open source - Pricing: open source - Self-hostable: yes - Docs: https://docs.searxng.org/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a web grounding component — what it knows. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the web grounding layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Exa | Embeddings-native web search built for agents, with full-content retrieval. | proprietary | no | | Tavily | Search API that returns model-ready answers rather than ten blue links. | proprietary | no | | Firecrawl | Turns any site into clean markdown, including the JavaScript-heavy ones. | open source | yes | | Brave Search API | Independent web index with straightforward per-query pricing. | proprietary | no | Head to head: - https://newagent.build/compare/searxng-vs-exa - https://newagent.build/compare/searxng-vs-tavily - https://newagent.build/compare/searxng-vs-firecrawl - https://newagent.build/compare/searxng-vs-brave-search ## Curated stacks using it - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?search=searxng&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # MCP (Model Context Protocol) MCP (Model Context Protocol) is a tools & integrations option for building AI agents. The open protocol for exposing tools and data to agents. The default answer. ## Facts - Layer: Tools & integrations (What it touches) - Vendor: Model Context Protocol (github.com/modelcontextprotocol/servers) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: ts, py - Install: `npm install @modelcontextprotocol/sdk` - Docs: https://modelcontextprotocol.io/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a tools & integrations component — what it touches. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing Write your integrations as MCP servers and every harness on this page can use them unchanged. ## Alternatives in the tools & integrations layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Composio | Managed auth and tool-calling for 250+ SaaS apps. | open source | no | | Arcade | Tool-calling with per-user OAuth, so the agent acts as the user, not as root. | open source | yes | | Browserbase | Managed headless browsers with stealth and session replay, plus Stagehand. | proprietary | no | | Playwright MCP | Drive a real local browser over MCP, using the accessibility tree. | open source | yes | | Nango | Open-source integration platform — you own the OAuth and the sync logic. | open source | yes | Head to head: - https://newagent.build/compare/mcp-vs-composio - https://newagent.build/compare/mcp-vs-arcade - https://newagent.build/compare/mcp-vs-browserbase - https://newagent.build/compare/mcp-vs-playwright-mcp - https://newagent.build/compare/mcp-vs-nango ## Curated stacks using it - Weekend prototype: Smallest thing that is still honest. Local, cheap, no accounts beyond one model key. - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?tools=mcp&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Composio Composio is a tools & integrations option for building AI agents. Managed auth and tool-calling for 250+ SaaS apps. ## Facts - Layer: Tools & integrations (What it touches) - Vendor: Composio (github.com/ComposioHQ/composio) - Licence: open source - Pricing: freemium - Self-hostable: no - SDKs: py, ts - Install: `pip install composio` - Requires: `COMPOSIO_API_KEY` - Docs: https://composio.dev/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a tools & integrations component — what it touches. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the tools & integrations layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | MCP (Model Context Protocol) | The open protocol for exposing tools and data to agents. The default answer. | open source | yes | | Arcade | Tool-calling with per-user OAuth, so the agent acts as the user, not as root. | open source | yes | | Browserbase | Managed headless browsers with stealth and session replay, plus Stagehand. | proprietary | no | | Playwright MCP | Drive a real local browser over MCP, using the accessibility tree. | open source | yes | | Nango | Open-source integration platform — you own the OAuth and the sync logic. | open source | yes | Head to head: - https://newagent.build/compare/composio-vs-mcp - https://newagent.build/compare/composio-vs-arcade - https://newagent.build/compare/composio-vs-browserbase - https://newagent.build/compare/composio-vs-playwright-mcp - https://newagent.build/compare/composio-vs-nango ## Curated stacks using it - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?tools=composio&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Arcade Arcade is a tools & integrations option for building AI agents. Tool-calling with per-user OAuth, so the agent acts as the user, not as root. ## Facts - Layer: Tools & integrations (What it touches) - Vendor: Arcade - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Requires: `ARCADE_API_KEY` - Docs: https://www.arcade.dev/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a tools & integrations component — what it touches. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Worth knowing The per-user token model is what stops one user's agent reading another user's mailbox. ## Alternatives in the tools & integrations layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | MCP (Model Context Protocol) | The open protocol for exposing tools and data to agents. The default answer. | open source | yes | | Composio | Managed auth and tool-calling for 250+ SaaS apps. | open source | no | | Browserbase | Managed headless browsers with stealth and session replay, plus Stagehand. | proprietary | no | | Playwright MCP | Drive a real local browser over MCP, using the accessibility tree. | open source | yes | | Nango | Open-source integration platform — you own the OAuth and the sync logic. | open source | yes | Head to head: - https://newagent.build/compare/arcade-vs-mcp - https://newagent.build/compare/arcade-vs-composio - https://newagent.build/compare/arcade-vs-browserbase - https://newagent.build/compare/arcade-vs-playwright-mcp - https://newagent.build/compare/arcade-vs-nango ## Curated stacks using it - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?tools=arcade&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Browserbase Browserbase is a tools & integrations option for building AI agents. Managed headless browsers with stealth and session replay, plus Stagehand. ## Facts - Layer: Tools & integrations (What it touches) - Vendor: Browserbase - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: ts, py - Requires: `BROWSERBASE_API_KEY`, `BROWSERBASE_PROJECT_ID` - Docs: https://www.browserbase.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a tools & integrations component — what it touches. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the tools & integrations layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | MCP (Model Context Protocol) | The open protocol for exposing tools and data to agents. The default answer. | open source | yes | | Composio | Managed auth and tool-calling for 250+ SaaS apps. | open source | no | | Arcade | Tool-calling with per-user OAuth, so the agent acts as the user, not as root. | open source | yes | | Playwright MCP | Drive a real local browser over MCP, using the accessibility tree. | open source | yes | | Nango | Open-source integration platform — you own the OAuth and the sync logic. | open source | yes | Head to head: - https://newagent.build/compare/browserbase-vs-mcp - https://newagent.build/compare/browserbase-vs-composio - https://newagent.build/compare/browserbase-vs-arcade - https://newagent.build/compare/browserbase-vs-playwright-mcp - https://newagent.build/compare/browserbase-vs-nango ## Curated stacks using it - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?tools=browserbase&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Playwright MCP Playwright MCP is a tools & integrations option for building AI agents. Drive a real local browser over MCP, using the accessibility tree. ## Facts - Layer: Tools & integrations (What it touches) - Vendor: Microsoft (github.com/data-privacy-stack/presidio) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: ts - Install: `npm install -g @playwright/mcp` - Docs: https://github.com/microsoft/playwright-mcp - Last verified: 2026-08-21 ## When to reach for it - You are choosing a tools & integrations component — what it touches. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the tools & integrations layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | MCP (Model Context Protocol) | The open protocol for exposing tools and data to agents. The default answer. | open source | yes | | Composio | Managed auth and tool-calling for 250+ SaaS apps. | open source | no | | Arcade | Tool-calling with per-user OAuth, so the agent acts as the user, not as root. | open source | yes | | Browserbase | Managed headless browsers with stealth and session replay, plus Stagehand. | proprietary | no | | Nango | Open-source integration platform — you own the OAuth and the sync logic. | open source | yes | Head to head: - https://newagent.build/compare/playwright-mcp-vs-mcp - https://newagent.build/compare/playwright-mcp-vs-composio - https://newagent.build/compare/playwright-mcp-vs-arcade - https://newagent.build/compare/playwright-mcp-vs-browserbase - https://newagent.build/compare/playwright-mcp-vs-nango ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?tools=playwright-mcp&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Nango Nango is a tools & integrations option for building AI agents. Open-source integration platform — you own the OAuth and the sync logic. ## Facts - Layer: Tools & integrations (What it touches) - Vendor: Nango (github.com/NangoHQ/nango) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: ts, py - Requires: `NANGO_SECRET_KEY` - Docs: https://www.nango.dev/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a tools & integrations component — what it touches. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Alternatives in the tools & integrations layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | MCP (Model Context Protocol) | The open protocol for exposing tools and data to agents. The default answer. | open source | yes | | Composio | Managed auth and tool-calling for 250+ SaaS apps. | open source | no | | Arcade | Tool-calling with per-user OAuth, so the agent acts as the user, not as root. | open source | yes | | Browserbase | Managed headless browsers with stealth and session replay, plus Stagehand. | proprietary | no | | Playwright MCP | Drive a real local browser over MCP, using the accessibility tree. | open source | yes | Head to head: - https://newagent.build/compare/nango-vs-mcp - https://newagent.build/compare/nango-vs-composio - https://newagent.build/compare/nango-vs-arcade - https://newagent.build/compare/nango-vs-browserbase - https://newagent.build/compare/nango-vs-playwright-mcp ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?tools=nango&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Llama Guard Llama Guard is a security option for building AI agents. Open-weight classifier for input and output safety. Runs wherever you run it. ## Facts - Layer: Security (What stops it) - Vendor: Meta (github.com/meta-llama/PurpleLlama) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py - Job within the layer: content - Docs: https://www.llama.com/docs/model-cards-and-prompt-formats/llama-guard-3/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. - Within security, it does one specific job: content filtering. ## Worth knowing Cheap first line: classify the user turn before it reaches the expensive model. ## Published benchmark scores - Agent Data Injection — attack success rate: 50% — peer-reviewed, Agent Data Injection Attacks, 2026-07. https://arxiv.org/pdf/2607.05120 - Agent Data Injection — detection rate: 34.9% — peer-reviewed, Agent Data Injection Attacks, 2026-07. https://arxiv.org/pdf/2607.05120 - Recall @ 1% FPR (English): 97.5% — self-reported, LlamaFirewall (Meta), 2026-05. https://arxiv.org/pdf/2505.03574 ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/llama-guard-vs-guardrails-ai - https://newagent.build/compare/llama-guard-vs-nemo-guardrails - https://newagent.build/compare/llama-guard-vs-lakera - https://newagent.build/compare/llama-guard-vs-rebuff - https://newagent.build/compare/llama-guard-vs-invariant - https://newagent.build/compare/llama-guard-vs-descope-agentic - https://newagent.build/compare/llama-guard-vs-auth0-ai - https://newagent.build/compare/llama-guard-vs-oso ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=llama-guard&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Guardrails AI Guardrails AI is a security option for building AI agents. Declarative input/output validators with a hub of prebuilt checks. ## Facts - Layer: Security (What stops it) - Vendor: Guardrails AI (github.com/guardrails-ai/guardrails) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py - Job within the layer: content - Install: `pip install guardrails-ai` - Docs: https://www.guardrailsai.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. - Within security, it does one specific job: content filtering. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/guardrails-ai-vs-llama-guard - https://newagent.build/compare/guardrails-ai-vs-nemo-guardrails - https://newagent.build/compare/guardrails-ai-vs-lakera - https://newagent.build/compare/guardrails-ai-vs-rebuff - https://newagent.build/compare/guardrails-ai-vs-invariant - https://newagent.build/compare/guardrails-ai-vs-descope-agentic - https://newagent.build/compare/guardrails-ai-vs-auth0-ai - https://newagent.build/compare/guardrails-ai-vs-oso ## Curated stacks using it - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=guardrails-ai&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # NeMo Guardrails NeMo Guardrails is a security option for building AI agents. NVIDIA's programmable rails — constrain topic, dialogue path and tool use. ## Facts - Layer: Security (What stops it) - Vendor: NVIDIA (github.com/NVIDIA-NeMo/Guardrails) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py - Job within the layer: policy - Install: `pip install nemoguardrails` - Docs: https://github.com/NVIDIA/NeMo-Guardrails - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. - Within security, it does one specific job: policy & topic control. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/nemo-guardrails-vs-invariant - https://newagent.build/compare/nemo-guardrails-vs-llama-guard - https://newagent.build/compare/nemo-guardrails-vs-guardrails-ai - https://newagent.build/compare/nemo-guardrails-vs-lakera - https://newagent.build/compare/nemo-guardrails-vs-rebuff - https://newagent.build/compare/nemo-guardrails-vs-descope-agentic - https://newagent.build/compare/nemo-guardrails-vs-auth0-ai - https://newagent.build/compare/nemo-guardrails-vs-oso ## Curated stacks using it - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=nemo-guardrails&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Lakera Guard Lakera Guard is a security option for building AI agents. Prompt-injection and jailbreak detection as a low-latency API call. ## Facts - Layer: Security (What stops it) - Vendor: Lakera - Licence: proprietary - Pricing: freemium - Self-hostable: no - Job within the layer: injection - Requires: `LAKERA_GUARD_API_KEY` - Docs: https://www.lakera.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You are happy to depend on a hosted service and would rather not operate this yourself. - Within security, it does one specific job: prompt-injection defence. ## Worth knowing Worth it the moment your agent reads text written by someone who is not your user. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/lakera-vs-rebuff - https://newagent.build/compare/lakera-vs-llama-guard - https://newagent.build/compare/lakera-vs-guardrails-ai - https://newagent.build/compare/lakera-vs-nemo-guardrails - https://newagent.build/compare/lakera-vs-invariant - https://newagent.build/compare/lakera-vs-descope-agentic - https://newagent.build/compare/lakera-vs-auth0-ai - https://newagent.build/compare/lakera-vs-oso ## Curated stacks using it - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=lakera&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Rebuff Rebuff is a security option for building AI agents. Open-source prompt-injection detector with canary tokens. ## Facts - Layer: Security (What stops it) - Vendor: Protect AI (github.com/protectai/rebuff) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py, ts - Job within the layer: injection - Install: `pip install rebuff` - Docs: https://github.com/protectai/rebuff - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. - Within security, it does one specific job: prompt-injection defence. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/rebuff-vs-lakera - https://newagent.build/compare/rebuff-vs-llama-guard - https://newagent.build/compare/rebuff-vs-guardrails-ai - https://newagent.build/compare/rebuff-vs-nemo-guardrails - https://newagent.build/compare/rebuff-vs-invariant - https://newagent.build/compare/rebuff-vs-descope-agentic - https://newagent.build/compare/rebuff-vs-auth0-ai - https://newagent.build/compare/rebuff-vs-oso ## Curated stacks using it - Weekend prototype: Smallest thing that is still honest. Local, cheap, no accounts beyond one model key. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=rebuff&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Invariant Invariant is a security option for building AI agents. Policy engine and analyzer for agent traces and MCP tool calls. ## Facts - Layer: Security (What stops it) - Vendor: Invariant Labs - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py - Job within the layer: policy - Docs: https://invariantlabs.ai/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - Within security, it does one specific job: policy & topic control. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/invariant-vs-nemo-guardrails - https://newagent.build/compare/invariant-vs-llama-guard - https://newagent.build/compare/invariant-vs-guardrails-ai - https://newagent.build/compare/invariant-vs-lakera - https://newagent.build/compare/invariant-vs-rebuff - https://newagent.build/compare/invariant-vs-descope-agentic - https://newagent.build/compare/invariant-vs-auth0-ai - https://newagent.build/compare/invariant-vs-oso ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=invariant&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Descope Agentic Identity Descope Agentic Identity is a security option for building AI agents. Identity, delegated consent and scoped tokens for agents acting on behalf of users. ## Facts - Layer: Security (What stops it) - Vendor: Descope - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: ts, py - Job within the layer: identity - Requires: `DESCOPE_PROJECT_ID`, `DESCOPE_MANAGEMENT_KEY` - Docs: https://www.descope.com/agentic-identity - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You are happy to depend on a hosted service and would rather not operate this yourself. - Within security, it does one specific job: agent identity & delegated auth. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/descope-agentic-vs-auth0-ai - https://newagent.build/compare/descope-agentic-vs-llama-guard - https://newagent.build/compare/descope-agentic-vs-guardrails-ai - https://newagent.build/compare/descope-agentic-vs-nemo-guardrails - https://newagent.build/compare/descope-agentic-vs-lakera - https://newagent.build/compare/descope-agentic-vs-rebuff - https://newagent.build/compare/descope-agentic-vs-invariant - https://newagent.build/compare/descope-agentic-vs-oso ## Curated stacks using it - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=descope-agentic&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Auth0 for AI Agents Auth0 for AI Agents is a security option for building AI agents. Token vault, async user approval and fine-grained authorization for tool calls. ## Facts - Layer: Security (What stops it) - Vendor: Auth0 - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: ts, py - Job within the layer: identity - Requires: `AUTH0_DOMAIN`, `AUTH0_CLIENT_ID`, `AUTH0_CLIENT_SECRET` - Docs: https://auth0.com/ai - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You are happy to depend on a hosted service and would rather not operate this yourself. - Within security, it does one specific job: agent identity & delegated auth. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/auth0-ai-vs-descope-agentic - https://newagent.build/compare/auth0-ai-vs-llama-guard - https://newagent.build/compare/auth0-ai-vs-guardrails-ai - https://newagent.build/compare/auth0-ai-vs-nemo-guardrails - https://newagent.build/compare/auth0-ai-vs-lakera - https://newagent.build/compare/auth0-ai-vs-rebuff - https://newagent.build/compare/auth0-ai-vs-invariant - https://newagent.build/compare/auth0-ai-vs-oso ## Curated stacks using it - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=auth0-ai&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Oso Oso is a security option for building AI agents. Authorization-as-a-service — decide what this agent may touch, per record. ## Facts - Layer: Security (What stops it) - Vendor: Oso - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Job within the layer: authz - Requires: `OSO_AUTH_API_KEY` - Docs: https://www.osohq.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - Within security, it does one specific job: authorization. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/oso-vs-llama-guard - https://newagent.build/compare/oso-vs-guardrails-ai - https://newagent.build/compare/oso-vs-nemo-guardrails - https://newagent.build/compare/oso-vs-lakera - https://newagent.build/compare/oso-vs-rebuff - https://newagent.build/compare/oso-vs-invariant - https://newagent.build/compare/oso-vs-descope-agentic - https://newagent.build/compare/oso-vs-auth0-ai ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=oso&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Microsoft Presidio Microsoft Presidio is a security option for building AI agents. PII detection and redaction before text ever reaches a model provider. ## Facts - Layer: Security (What stops it) - Vendor: Microsoft (github.com/data-privacy-stack/presidio) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py - Job within the layer: pii - Install: `pip install presidio-analyzer presidio-anonymizer` - Docs: https://microsoft.github.io/presidio/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. - Within security, it does one specific job: pii detection & redaction. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/presidio-vs-llama-guard - https://newagent.build/compare/presidio-vs-guardrails-ai - https://newagent.build/compare/presidio-vs-nemo-guardrails - https://newagent.build/compare/presidio-vs-lakera - https://newagent.build/compare/presidio-vs-rebuff - https://newagent.build/compare/presidio-vs-invariant - https://newagent.build/compare/presidio-vs-descope-agentic - https://newagent.build/compare/presidio-vs-auth0-ai ## Curated stacks using it - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=presidio&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Cloudflare AI Gateway Cloudflare AI Gateway is a security option for building AI agents. Rate limiting, caching, spend caps and logging in front of any provider. ## Facts - Layer: Security (What stops it) - Vendor: Cloudflare - Licence: proprietary - Pricing: freemium - Self-hostable: no - Job within the layer: gateway - Requires: `CLOUDFLARE_ACCOUNT_ID`, `CLOUDFLARE_GATEWAY_ID` - Docs: https://developers.cloudflare.com/ai-gateway/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You are happy to depend on a hosted service and would rather not operate this yourself. - Within security, it does one specific job: gateway & spend control. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/cf-ai-gateway-vs-llama-guard - https://newagent.build/compare/cf-ai-gateway-vs-guardrails-ai - https://newagent.build/compare/cf-ai-gateway-vs-nemo-guardrails - https://newagent.build/compare/cf-ai-gateway-vs-lakera - https://newagent.build/compare/cf-ai-gateway-vs-rebuff - https://newagent.build/compare/cf-ai-gateway-vs-invariant - https://newagent.build/compare/cf-ai-gateway-vs-descope-agentic - https://newagent.build/compare/cf-ai-gateway-vs-auth0-ai ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=cf-ai-gateway&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # iron-proxy iron-proxy is a security option for building AI agents. Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. ## Facts - Layer: Security (What stops it) - Vendor: iron.sh (github.com/paradigmxyz/iron-proxy) - Licence: open source - Pricing: open source - Self-hostable: yes - Job within the layer: egress - Docs: https://github.com/paradigmxyz/iron-proxy - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. - Within security, it does one specific job: egress control. ## Worth knowing Single Go binary, single YAML allowlist. Denies cloud metadata (169.254.169.254) and loopback by default, closing the SSRF and DNS-rebinding path. Hermes Agent wires this in as `hermes egress`. Now maintained by Paradigm, who also build Centaur. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | Head to head: - https://newagent.build/compare/iron-proxy-vs-llama-guard - https://newagent.build/compare/iron-proxy-vs-guardrails-ai - https://newagent.build/compare/iron-proxy-vs-nemo-guardrails - https://newagent.build/compare/iron-proxy-vs-lakera - https://newagent.build/compare/iron-proxy-vs-rebuff - https://newagent.build/compare/iron-proxy-vs-invariant - https://newagent.build/compare/iron-proxy-vs-descope-agentic - https://newagent.build/compare/iron-proxy-vs-auth0-ai ## Curated stacks using it - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=iron-proxy&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # gVisor gVisor is a security option for building AI agents. Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. ## Facts - Layer: Security (What stops it) - Vendor: gVisor (github.com/google/gvisor) - Licence: open source - Pricing: open source - Self-hostable: yes - Job within the layer: sandbox - Install: `# runsc — see https://gvisor.dev/docs/user_guide/install/` - Docs: https://gvisor.dev/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. - Within security, it does one specific job: sandboxing. ## Worth knowing Fast to start, but 10-30% overhead on I/O-heavy work. Good for compute-bound agent code; reach for a microVM if the threat model is serious. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | Head to head: - https://newagent.build/compare/gvisor-vs-firecracker - https://newagent.build/compare/gvisor-vs-kata - https://newagent.build/compare/gvisor-vs-host-only - https://newagent.build/compare/gvisor-vs-e2b - https://newagent.build/compare/gvisor-vs-llama-guard - https://newagent.build/compare/gvisor-vs-guardrails-ai - https://newagent.build/compare/gvisor-vs-nemo-guardrails - https://newagent.build/compare/gvisor-vs-lakera ## Curated stacks using it - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=gvisor&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Firecracker Firecracker is a security option for building AI agents. AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. ## Facts - Layer: Security (What stops it) - Vendor: Amazon Web Services (github.com/firecracker-microvm/firecracker) - Licence: open source - Pricing: open source - Self-hostable: yes - Job within the layer: sandbox - Docs: https://firecracker-microvm.github.io/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. - Within security, it does one specific job: sandboxing. ## Worth knowing The default answer for running code a model wrote. Under 5 MiB overhead per VM, so per-task disposable sandboxes are affordable. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | Head to head: - https://newagent.build/compare/firecracker-vs-gvisor - https://newagent.build/compare/firecracker-vs-kata - https://newagent.build/compare/firecracker-vs-host-only - https://newagent.build/compare/firecracker-vs-e2b - https://newagent.build/compare/firecracker-vs-llama-guard - https://newagent.build/compare/firecracker-vs-guardrails-ai - https://newagent.build/compare/firecracker-vs-nemo-guardrails - https://newagent.build/compare/firecracker-vs-lakera ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=firecracker&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Kata Containers Kata Containers is a security option for building AI agents. MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. ## Facts - Layer: Security (What stops it) - Vendor: Kata Containers (github.com/kata-containers/kata-containers) - Licence: open source - Pricing: open source - Self-hostable: yes - Job within the layer: sandbox - Docs: https://katacontainers.io/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. - Within security, it does one specific job: sandboxing. ## Worth knowing The least disruptive option if you already run Kubernetes: VM-grade isolation without changing how workloads are packaged. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | Head to head: - https://newagent.build/compare/kata-vs-gvisor - https://newagent.build/compare/kata-vs-firecracker - https://newagent.build/compare/kata-vs-host-only - https://newagent.build/compare/kata-vs-e2b - https://newagent.build/compare/kata-vs-llama-guard - https://newagent.build/compare/kata-vs-guardrails-ai - https://newagent.build/compare/kata-vs-nemo-guardrails - https://newagent.build/compare/kata-vs-lakera ## Curated stacks using it - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=kata&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # No isolation (host process) No isolation (host process) is a security option for building AI agents. The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. ## Facts - Layer: Security (What stops it) - Vendor: OWASP - Licence: open source - Pricing: open source - Self-hostable: yes - Job within the layer: sandbox - Docs: https://owasp.org/www-project-top-10-for-large-language-model-applications/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. - Within security, it does one specific job: sandboxing. ## Worth knowing A container is not a sandbox. If the agent runs model-written code, this is not a defensible choice. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | E2B | Firecracker sandboxes for running code the agent wrote, safely. | open source | yes | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | Head to head: - https://newagent.build/compare/host-only-vs-gvisor - https://newagent.build/compare/host-only-vs-firecracker - https://newagent.build/compare/host-only-vs-kata - https://newagent.build/compare/host-only-vs-e2b - https://newagent.build/compare/host-only-vs-llama-guard - https://newagent.build/compare/host-only-vs-guardrails-ai - https://newagent.build/compare/host-only-vs-nemo-guardrails - https://newagent.build/compare/host-only-vs-lakera ## Curated stacks using it - Weekend prototype: Smallest thing that is still honest. Local, cheap, no accounts beyond one model key. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=host-only&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Langfuse Langfuse is a observability & evals option for building AI agents. Open-source tracing, prompt management and evals. Self-hostable in one compose file. ## Facts - Layer: Observability & evals (How you run it) - Vendor: Langfuse (github.com/langfuse/langfuse) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Install: `pip install langfuse` - Requires: `LANGFUSE_PUBLIC_KEY`, `LANGFUSE_SECRET_KEY` - Docs: https://langfuse.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a observability & evals component — how you run it. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Alternatives in the observability & evals layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | LangSmith | Tracing, datasets and evals from the LangChain team. | proprietary | no | | Braintrust | Eval-first platform — scorers, datasets and a playground for prompt iteration. | proprietary | yes | | Arize Phoenix | OpenTelemetry-native tracing and evals you can run locally. | open source | yes | | Pydantic Logfire | OpenTelemetry observability with first-class Python and Pydantic AI support. | open source | no | | promptfoo | Local eval and red-team harness that runs in CI. No account needed. | open source | yes | Head to head: - https://newagent.build/compare/langfuse-vs-langsmith - https://newagent.build/compare/langfuse-vs-braintrust - https://newagent.build/compare/langfuse-vs-phoenix - https://newagent.build/compare/langfuse-vs-logfire - https://newagent.build/compare/langfuse-vs-promptfoo ## Curated stacks using it - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?observability=langfuse&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # LangSmith LangSmith is a observability & evals option for building AI agents. Tracing, datasets and evals from the LangChain team. ## Facts - Layer: Observability & evals (How you run it) - Vendor: LangChain (github.com/langchain-ai/langgraph) - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: py, ts - Requires: `LANGSMITH_API_KEY` - Docs: https://www.langchain.com/langsmith - Last verified: 2026-08-21 ## When to reach for it - You are choosing a observability & evals component — how you run it. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the observability & evals layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Langfuse | Open-source tracing, prompt management and evals. Self-hostable in one compose file. | open source | yes | | Braintrust | Eval-first platform — scorers, datasets and a playground for prompt iteration. | proprietary | yes | | Arize Phoenix | OpenTelemetry-native tracing and evals you can run locally. | open source | yes | | Pydantic Logfire | OpenTelemetry observability with first-class Python and Pydantic AI support. | open source | no | | promptfoo | Local eval and red-team harness that runs in CI. No account needed. | open source | yes | Head to head: - https://newagent.build/compare/langsmith-vs-langfuse - https://newagent.build/compare/langsmith-vs-braintrust - https://newagent.build/compare/langsmith-vs-phoenix - https://newagent.build/compare/langsmith-vs-logfire - https://newagent.build/compare/langsmith-vs-promptfoo ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?observability=langsmith&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Braintrust Braintrust is a observability & evals option for building AI agents. Eval-first platform — scorers, datasets and a playground for prompt iteration. ## Facts - Layer: Observability & evals (How you run it) - Vendor: Braintrust - Licence: proprietary - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Requires: `BRAINTRUST_API_KEY` - Docs: https://www.braintrust.dev/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a observability & evals component — how you run it. ## Alternatives in the observability & evals layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Langfuse | Open-source tracing, prompt management and evals. Self-hostable in one compose file. | open source | yes | | LangSmith | Tracing, datasets and evals from the LangChain team. | proprietary | no | | Arize Phoenix | OpenTelemetry-native tracing and evals you can run locally. | open source | yes | | Pydantic Logfire | OpenTelemetry observability with first-class Python and Pydantic AI support. | open source | no | | promptfoo | Local eval and red-team harness that runs in CI. No account needed. | open source | yes | Head to head: - https://newagent.build/compare/braintrust-vs-langfuse - https://newagent.build/compare/braintrust-vs-langsmith - https://newagent.build/compare/braintrust-vs-phoenix - https://newagent.build/compare/braintrust-vs-logfire - https://newagent.build/compare/braintrust-vs-promptfoo ## Curated stacks using it - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?observability=braintrust&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Arize Phoenix Arize Phoenix is a observability & evals option for building AI agents. OpenTelemetry-native tracing and evals you can run locally. ## Facts - Layer: Observability & evals (How you run it) - Vendor: Arize (github.com/Arize-ai/phoenix) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: py, ts - Install: `pip install arize-phoenix` - Docs: https://phoenix.arize.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a observability & evals component — how you run it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the observability & evals layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Langfuse | Open-source tracing, prompt management and evals. Self-hostable in one compose file. | open source | yes | | LangSmith | Tracing, datasets and evals from the LangChain team. | proprietary | no | | Braintrust | Eval-first platform — scorers, datasets and a playground for prompt iteration. | proprietary | yes | | Pydantic Logfire | OpenTelemetry observability with first-class Python and Pydantic AI support. | open source | no | | promptfoo | Local eval and red-team harness that runs in CI. No account needed. | open source | yes | Head to head: - https://newagent.build/compare/phoenix-vs-langfuse - https://newagent.build/compare/phoenix-vs-langsmith - https://newagent.build/compare/phoenix-vs-braintrust - https://newagent.build/compare/phoenix-vs-logfire - https://newagent.build/compare/phoenix-vs-promptfoo ## Curated stacks using it - Weekend prototype: Smallest thing that is still honest. Local, cheap, no accounts beyond one model key. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?observability=phoenix&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Pydantic Logfire Pydantic Logfire is a observability & evals option for building AI agents. OpenTelemetry observability with first-class Python and Pydantic AI support. ## Facts - Layer: Observability & evals (How you run it) - Vendor: Pydantic (github.com/pydantic/pydantic-ai) - Licence: open source - Pricing: freemium - Self-hostable: no - SDKs: py, ts - Install: `pip install logfire` - Requires: `LOGFIRE_TOKEN` - Docs: https://pydantic.dev/logfire - Last verified: 2026-08-21 ## When to reach for it - You are choosing a observability & evals component — how you run it. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the observability & evals layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Langfuse | Open-source tracing, prompt management and evals. Self-hostable in one compose file. | open source | yes | | LangSmith | Tracing, datasets and evals from the LangChain team. | proprietary | no | | Braintrust | Eval-first platform — scorers, datasets and a playground for prompt iteration. | proprietary | yes | | Arize Phoenix | OpenTelemetry-native tracing and evals you can run locally. | open source | yes | | promptfoo | Local eval and red-team harness that runs in CI. No account needed. | open source | yes | Head to head: - https://newagent.build/compare/logfire-vs-langfuse - https://newagent.build/compare/logfire-vs-langsmith - https://newagent.build/compare/logfire-vs-braintrust - https://newagent.build/compare/logfire-vs-phoenix - https://newagent.build/compare/logfire-vs-promptfoo ## Curated stacks using it - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?observability=logfire&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # promptfoo promptfoo is a observability & evals option for building AI agents. Local eval and red-team harness that runs in CI. No account needed. ## Facts - Layer: Observability & evals (How you run it) - Vendor: promptfoo (github.com/promptfoo/promptfoo) - Licence: open source - Pricing: open source - Self-hostable: yes - SDKs: ts, py - Install: `npm install -g promptfoo` - Docs: https://www.promptfoo.dev/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a observability & evals component — how you run it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing Its red-team mode doubles as your prompt-injection regression suite. ## Alternatives in the observability & evals layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Langfuse | Open-source tracing, prompt management and evals. Self-hostable in one compose file. | open source | yes | | LangSmith | Tracing, datasets and evals from the LangChain team. | proprietary | no | | Braintrust | Eval-first platform — scorers, datasets and a playground for prompt iteration. | proprietary | yes | | Arize Phoenix | OpenTelemetry-native tracing and evals you can run locally. | open source | yes | | Pydantic Logfire | OpenTelemetry observability with first-class Python and Pydantic AI support. | open source | no | Head to head: - https://newagent.build/compare/promptfoo-vs-langfuse - https://newagent.build/compare/promptfoo-vs-langsmith - https://newagent.build/compare/promptfoo-vs-braintrust - https://newagent.build/compare/promptfoo-vs-phoenix - https://newagent.build/compare/promptfoo-vs-logfire ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?observability=promptfoo&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Docker Compose Docker Compose is a runtime & deploy option for building AI agents. One file, one box, everything local. The honest starting point. ## Facts - Layer: Runtime & deploy (How you run it) - Vendor: Docker - Licence: open source - Pricing: open source - Self-hostable: yes - Docs: https://docs.docker.com/compose/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a runtime & deploy component — how you run it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing The generated docker-compose.yml below is runnable as-is. ## Alternatives in the runtime & deploy layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Vercel | Streaming-friendly serverless for TypeScript agents with a web front end. | proprietary | no | | Cloudflare Workers | Edge runtime with Durable Objects for long-lived, stateful agent sessions. | proprietary | no | | Modal | Python-native serverless with GPUs and sandboxes, scale to zero. | proprietary | no | | Fly.io | Long-running containers close to users, without owning a Kubernetes cluster. | proprietary | no | | Temporal | Durable execution — the run survives a crash, a deploy and a week of waiting. | open source | yes | | Centaur | Paradigm's self-hosted control plane: one shared team agent, an isolated k8s sandbox per Slack thread, and no raw API keys in the agent's environment. | open source | yes | | Kubernetes | Your existing cluster. Nothing new for the platform team to approve. | open source | yes | Head to head: - https://newagent.build/compare/docker-compose-vs-vercel - https://newagent.build/compare/docker-compose-vs-cloudflare-workers - https://newagent.build/compare/docker-compose-vs-modal - https://newagent.build/compare/docker-compose-vs-fly - https://newagent.build/compare/docker-compose-vs-temporal - https://newagent.build/compare/docker-compose-vs-centaur - https://newagent.build/compare/docker-compose-vs-k8s ## Curated stacks using it - Weekend prototype: Smallest thing that is still honest. Local, cheap, no accounts beyond one model key. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?runtime=docker-compose&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Vercel Vercel is a runtime & deploy option for building AI agents. Streaming-friendly serverless for TypeScript agents with a web front end. ## Facts - Layer: Runtime & deploy (How you run it) - Vendor: Vercel (github.com/vercel/ai) - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: ts - Install: `npm i -g vercel` - Docs: https://vercel.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a runtime & deploy component — how you run it. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the runtime & deploy layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Docker Compose | One file, one box, everything local. The honest starting point. | open source | yes | | Cloudflare Workers | Edge runtime with Durable Objects for long-lived, stateful agent sessions. | proprietary | no | | Modal | Python-native serverless with GPUs and sandboxes, scale to zero. | proprietary | no | | Fly.io | Long-running containers close to users, without owning a Kubernetes cluster. | proprietary | no | | Temporal | Durable execution — the run survives a crash, a deploy and a week of waiting. | open source | yes | | Centaur | Paradigm's self-hosted control plane: one shared team agent, an isolated k8s sandbox per Slack thread, and no raw API keys in the agent's environment. | open source | yes | | Kubernetes | Your existing cluster. Nothing new for the platform team to approve. | open source | yes | Head to head: - https://newagent.build/compare/vercel-vs-docker-compose - https://newagent.build/compare/vercel-vs-cloudflare-workers - https://newagent.build/compare/vercel-vs-modal - https://newagent.build/compare/vercel-vs-fly - https://newagent.build/compare/vercel-vs-temporal - https://newagent.build/compare/vercel-vs-centaur - https://newagent.build/compare/vercel-vs-k8s ## Curated stacks using it - Customer-facing support agent: Talks to strangers, remembers them, and is assumed hostile-adjacent by default. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?runtime=vercel&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Cloudflare Workers Cloudflare Workers is a runtime & deploy option for building AI agents. Edge runtime with Durable Objects for long-lived, stateful agent sessions. ## Facts - Layer: Runtime & deploy (How you run it) - Vendor: Cloudflare - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: ts - Install: `npm i -g wrangler` - Docs: https://developers.cloudflare.com/agents/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a runtime & deploy component — how you run it. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the runtime & deploy layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Docker Compose | One file, one box, everything local. The honest starting point. | open source | yes | | Vercel | Streaming-friendly serverless for TypeScript agents with a web front end. | proprietary | no | | Modal | Python-native serverless with GPUs and sandboxes, scale to zero. | proprietary | no | | Fly.io | Long-running containers close to users, without owning a Kubernetes cluster. | proprietary | no | | Temporal | Durable execution — the run survives a crash, a deploy and a week of waiting. | open source | yes | | Centaur | Paradigm's self-hosted control plane: one shared team agent, an isolated k8s sandbox per Slack thread, and no raw API keys in the agent's environment. | open source | yes | | Kubernetes | Your existing cluster. Nothing new for the platform team to approve. | open source | yes | Head to head: - https://newagent.build/compare/cloudflare-workers-vs-docker-compose - https://newagent.build/compare/cloudflare-workers-vs-vercel - https://newagent.build/compare/cloudflare-workers-vs-modal - https://newagent.build/compare/cloudflare-workers-vs-fly - https://newagent.build/compare/cloudflare-workers-vs-temporal - https://newagent.build/compare/cloudflare-workers-vs-centaur - https://newagent.build/compare/cloudflare-workers-vs-k8s ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?runtime=cloudflare-workers&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Modal Modal is a runtime & deploy option for building AI agents. Python-native serverless with GPUs and sandboxes, scale to zero. ## Facts - Layer: Runtime & deploy (How you run it) - Vendor: Modal - Licence: proprietary - Pricing: freemium - Self-hostable: no - SDKs: py - Install: `pip install modal` - Requires: `MODAL_TOKEN_ID`, `MODAL_TOKEN_SECRET` - Docs: https://modal.com/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a runtime & deploy component — how you run it. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the runtime & deploy layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Docker Compose | One file, one box, everything local. The honest starting point. | open source | yes | | Vercel | Streaming-friendly serverless for TypeScript agents with a web front end. | proprietary | no | | Cloudflare Workers | Edge runtime with Durable Objects for long-lived, stateful agent sessions. | proprietary | no | | Fly.io | Long-running containers close to users, without owning a Kubernetes cluster. | proprietary | no | | Temporal | Durable execution — the run survives a crash, a deploy and a week of waiting. | open source | yes | | Centaur | Paradigm's self-hosted control plane: one shared team agent, an isolated k8s sandbox per Slack thread, and no raw API keys in the agent's environment. | open source | yes | | Kubernetes | Your existing cluster. Nothing new for the platform team to approve. | open source | yes | Head to head: - https://newagent.build/compare/modal-vs-docker-compose - https://newagent.build/compare/modal-vs-vercel - https://newagent.build/compare/modal-vs-cloudflare-workers - https://newagent.build/compare/modal-vs-fly - https://newagent.build/compare/modal-vs-temporal - https://newagent.build/compare/modal-vs-centaur - https://newagent.build/compare/modal-vs-k8s ## Curated stacks using it - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?runtime=modal&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Fly.io Fly.io is a runtime & deploy option for building AI agents. Long-running containers close to users, without owning a Kubernetes cluster. ## Facts - Layer: Runtime & deploy (How you run it) - Vendor: Fly.io - Licence: proprietary - Pricing: usage based - Self-hostable: no - Install: `brew install flyctl` - Docs: https://fly.io/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a runtime & deploy component — how you run it. - You are happy to depend on a hosted service and would rather not operate this yourself. ## Alternatives in the runtime & deploy layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Docker Compose | One file, one box, everything local. The honest starting point. | open source | yes | | Vercel | Streaming-friendly serverless for TypeScript agents with a web front end. | proprietary | no | | Cloudflare Workers | Edge runtime with Durable Objects for long-lived, stateful agent sessions. | proprietary | no | | Modal | Python-native serverless with GPUs and sandboxes, scale to zero. | proprietary | no | | Temporal | Durable execution — the run survives a crash, a deploy and a week of waiting. | open source | yes | | Centaur | Paradigm's self-hosted control plane: one shared team agent, an isolated k8s sandbox per Slack thread, and no raw API keys in the agent's environment. | open source | yes | | Kubernetes | Your existing cluster. Nothing new for the platform team to approve. | open source | yes | Head to head: - https://newagent.build/compare/fly-vs-docker-compose - https://newagent.build/compare/fly-vs-vercel - https://newagent.build/compare/fly-vs-cloudflare-workers - https://newagent.build/compare/fly-vs-modal - https://newagent.build/compare/fly-vs-temporal - https://newagent.build/compare/fly-vs-centaur - https://newagent.build/compare/fly-vs-k8s ## Curated stacks using it - Browser automation: Drives real web apps that have no API, on behalf of a specific signed-in user. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?runtime=fly&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # E2B E2B is a security option for building AI agents. Firecracker sandboxes for running code the agent wrote, safely. ## Facts - Layer: Security (What stops it) - Vendor: E2B (github.com/e2b-dev/E2B) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: py, ts - Job within the layer: sandbox - Install: `npm install @e2b/code-interpreter` - Requires: `E2B_API_KEY` - Docs: https://e2b.dev/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a security component — what stops it. - You need to run it on your own infrastructure, or your procurement process requires source access. - Within security, it does one specific job: sandboxing. ## Worth knowing If the agent writes code that then runs, this layer is not optional. ## Alternatives in the security layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | gVisor | Google's user-space kernel. Intercepts syscalls before they reach the host, without paying for a full VM. | open source | yes | | Firecracker | AWS microVMs — a separate Linux kernel per sandbox, booting in about 125ms. | open source | yes | | Kata Containers | MicroVM isolation behind the ordinary container API — drops into Kubernetes as a RuntimeClass. | open source | yes | | No isolation (host process) | The agent runs as an ordinary process on the host. Honest default only when it executes no code and reads nothing untrusted. | open source | yes | | Llama Guard | Open-weight classifier for input and output safety. Runs wherever you run it. | open source | yes | | Guardrails AI | Declarative input/output validators with a hub of prebuilt checks. | open source | yes | | NeMo Guardrails | NVIDIA's programmable rails — constrain topic, dialogue path and tool use. | open source | yes | | Lakera Guard | Prompt-injection and jailbreak detection as a low-latency API call. | proprietary | no | | Rebuff | Open-source prompt-injection detector with canary tokens. | open source | yes | | Invariant | Policy engine and analyzer for agent traces and MCP tool calls. | open source | yes | | Descope Agentic Identity | Identity, delegated consent and scoped tokens for agents acting on behalf of users. | proprietary | no | | Auth0 for AI Agents | Token vault, async user approval and fine-grained authorization for tool calls. | proprietary | no | | Oso | Authorization-as-a-service — decide what this agent may touch, per record. | open source | yes | | Microsoft Presidio | PII detection and redaction before text ever reaches a model provider. | open source | yes | | Cloudflare AI Gateway | Rate limiting, caching, spend caps and logging in front of any provider. | proprietary | no | | iron-proxy | Default-deny egress firewall for untrusted workloads. The sandbox holds worthless proxy tokens; real credentials are swapped in at the boundary. | open source | yes | Head to head: - https://newagent.build/compare/e2b-vs-gvisor - https://newagent.build/compare/e2b-vs-firecracker - https://newagent.build/compare/e2b-vs-kata - https://newagent.build/compare/e2b-vs-host-only - https://newagent.build/compare/e2b-vs-llama-guard - https://newagent.build/compare/e2b-vs-guardrails-ai - https://newagent.build/compare/e2b-vs-nemo-guardrails - https://newagent.build/compare/e2b-vs-lakera ## Curated stacks using it - Research analyst: Reads the whole internet, writes something with citations, runs code to check itself. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?security=e2b&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Temporal Temporal is a runtime & deploy option for building AI agents. Durable execution — the run survives a crash, a deploy and a week of waiting. ## Facts - Layer: Runtime & deploy (How you run it) - Vendor: Temporal (github.com/temporalio/temporal) - Licence: open source - Pricing: freemium - Self-hostable: yes - SDKs: ts, py - Docs: https://temporal.io/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a runtime & deploy component — how you run it. - You need to run it on your own infrastructure, or your procurement process requires source access. ## Alternatives in the runtime & deploy layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Docker Compose | One file, one box, everything local. The honest starting point. | open source | yes | | Vercel | Streaming-friendly serverless for TypeScript agents with a web front end. | proprietary | no | | Cloudflare Workers | Edge runtime with Durable Objects for long-lived, stateful agent sessions. | proprietary | no | | Modal | Python-native serverless with GPUs and sandboxes, scale to zero. | proprietary | no | | Fly.io | Long-running containers close to users, without owning a Kubernetes cluster. | proprietary | no | | Centaur | Paradigm's self-hosted control plane: one shared team agent, an isolated k8s sandbox per Slack thread, and no raw API keys in the agent's environment. | open source | yes | | Kubernetes | Your existing cluster. Nothing new for the platform team to approve. | open source | yes | Head to head: - https://newagent.build/compare/temporal-vs-docker-compose - https://newagent.build/compare/temporal-vs-vercel - https://newagent.build/compare/temporal-vs-cloudflare-workers - https://newagent.build/compare/temporal-vs-modal - https://newagent.build/compare/temporal-vs-fly - https://newagent.build/compare/temporal-vs-centaur - https://newagent.build/compare/temporal-vs-k8s ## Curated stacks using it - Long-running operations agent: Runs for hours, survives deploys, and stops to ask a human before it spends money. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?runtime=temporal&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Centaur Centaur is a runtime & deploy option for building AI agents. Paradigm's self-hosted control plane: one shared team agent, an isolated k8s sandbox per Slack thread, and no raw API keys in the agent's environment. ## Facts - Layer: Runtime & deploy (How you run it) - Vendor: Paradigm (github.com/paradigmxyz/centaur) - Licence: open source - Pricing: open source - Self-hostable: yes - Docs: https://github.com/paradigmxyz/centaur - Last verified: 2026-08-21 ## When to reach for it - You are choosing a runtime & deploy component — how you run it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Worth knowing Brings your own harness — it runs Claude Code, Codex or Amp inside the sandbox rather than replacing them. Credential boundaries mean the agent calls approved services without ever holding the key. ## Alternatives in the runtime & deploy layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Docker Compose | One file, one box, everything local. The honest starting point. | open source | yes | | Vercel | Streaming-friendly serverless for TypeScript agents with a web front end. | proprietary | no | | Cloudflare Workers | Edge runtime with Durable Objects for long-lived, stateful agent sessions. | proprietary | no | | Modal | Python-native serverless with GPUs and sandboxes, scale to zero. | proprietary | no | | Fly.io | Long-running containers close to users, without owning a Kubernetes cluster. | proprietary | no | | Temporal | Durable execution — the run survives a crash, a deploy and a week of waiting. | open source | yes | | Kubernetes | Your existing cluster. Nothing new for the platform team to approve. | open source | yes | Head to head: - https://newagent.build/compare/centaur-vs-docker-compose - https://newagent.build/compare/centaur-vs-vercel - https://newagent.build/compare/centaur-vs-cloudflare-workers - https://newagent.build/compare/centaur-vs-modal - https://newagent.build/compare/centaur-vs-fly - https://newagent.build/compare/centaur-vs-temporal - https://newagent.build/compare/centaur-vs-k8s ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?runtime=centaur&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Kubernetes Kubernetes is a runtime & deploy option for building AI agents. Your existing cluster. Nothing new for the platform team to approve. ## Facts - Layer: Runtime & deploy (How you run it) - Vendor: Kubernetes (github.com/kubernetes/kubernetes) - Licence: open source - Pricing: open source - Self-hostable: yes - Docs: https://kubernetes.io/ - Last verified: 2026-08-21 ## When to reach for it - You are choosing a runtime & deploy component — how you run it. - You need to run it on your own infrastructure, or your procurement process requires source access. - You want to start without a commercial conversation. ## Alternatives in the runtime & deploy layer | Option | What it does | Licence | Self-host | | --- | --- | --- | --- | | Docker Compose | One file, one box, everything local. The honest starting point. | open source | yes | | Vercel | Streaming-friendly serverless for TypeScript agents with a web front end. | proprietary | no | | Cloudflare Workers | Edge runtime with Durable Objects for long-lived, stateful agent sessions. | proprietary | no | | Modal | Python-native serverless with GPUs and sandboxes, scale to zero. | proprietary | no | | Fly.io | Long-running containers close to users, without owning a Kubernetes cluster. | proprietary | no | | Temporal | Durable execution — the run survives a crash, a deploy and a week of waiting. | open source | yes | | Centaur | Paradigm's self-hosted control plane: one shared team agent, an isolated k8s sandbox per Slack thread, and no raw API keys in the agent's environment. | open source | yes | Head to head: - https://newagent.build/compare/k8s-vs-docker-compose - https://newagent.build/compare/k8s-vs-vercel - https://newagent.build/compare/k8s-vs-cloudflare-workers - https://newagent.build/compare/k8s-vs-modal - https://newagent.build/compare/k8s-vs-fly - https://newagent.build/compare/k8s-vs-temporal - https://newagent.build/compare/k8s-vs-centaur ## Curated stacks using it - Regulated / on-prem: Nothing leaves the estate. Every component runs on infrastructure you control. ## Add it to a stack ```bash curl -fsSL "https://newagent.build/api/stack?runtime=k8s&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Curated stacks # Curated AI agent stacks ## Weekend prototype Smallest thing that is still honest. Local, cheap, no accounts beyond one model key. - Harness: Claude Code - Model provider: Anthropic - Memory: Postgres (rolled by hand) - Retrieval: Chroma - Web grounding: Tavily - Tools & integrations: MCP (Model Context Protocol) - Security: Rebuff + No isolation (host process) - Observability & evals: Arize Phoenix - Runtime & deploy: Docker Compose ```bash curl -fsSL "https://newagent.build/api/stack?harness=claude-code&model=anthropic&memory=postgres-memory&retrieval=chroma&search=tavily&tools=mcp&security=rebuff&security=host-only&observability=phoenix&runtime=docker-compose&format=sh" | sh ``` ## Customer-facing support agent Talks to strangers, remembers them, and is assumed hostile-adjacent by default. - Harness: Claude Agent SDK - Model provider: Anthropic - Framework: Vercel AI SDK - Memory: Zep - Retrieval: turbopuffer - Web grounding: Exa - Tools & integrations: Composio - Security: Lakera Guard + Microsoft Presidio + iron-proxy - Observability & evals: Langfuse - Runtime & deploy: Vercel ```bash curl -fsSL "https://newagent.build/api/stack?harness=claude-agent-sdk&model=anthropic&framework=vercel-ai-sdk&memory=zep&retrieval=turbopuffer&search=exa&tools=composio&security=lakera&security=presidio&security=iron-proxy&observability=langfuse&runtime=vercel&format=sh" | sh ``` ## Regulated / on-prem Nothing leaves the estate. Every component runs on infrastructure you control. - Harness: Letta - Model provider: vLLM - Framework: LangGraph - Memory: Postgres (rolled by hand) - Retrieval: pgvector - Web grounding: SearXNG - Tools & integrations: MCP (Model Context Protocol) - Security: Microsoft Presidio + NeMo Guardrails + Kata Containers + iron-proxy - Observability & evals: Langfuse - Runtime & deploy: Kubernetes ```bash curl -fsSL "https://newagent.build/api/stack?harness=letta&model=vllm&framework=langgraph&memory=postgres-memory&retrieval=pgvector&search=searxng&tools=mcp&security=presidio&security=nemo-guardrails&security=kata&security=iron-proxy&observability=langfuse&runtime=k8s&format=sh" | sh ``` ## Long-running operations agent Runs for hours, survives deploys, and stops to ask a human before it spends money. - Harness: Claude Agent SDK - Model provider: Anthropic - Framework: LangGraph - Memory: Mem0 - Retrieval: Qdrant - Web grounding: Exa - Tools & integrations: Arcade - Security: Auth0 for AI Agents + Lakera Guard + iron-proxy - Observability & evals: Braintrust - Runtime & deploy: Temporal ```bash curl -fsSL "https://newagent.build/api/stack?harness=claude-agent-sdk&model=anthropic&framework=langgraph&memory=mem0&retrieval=qdrant&search=exa&tools=arcade&security=auth0-ai&security=lakera&security=iron-proxy&observability=braintrust&runtime=temporal&format=sh" | sh ``` ## Research analyst Reads the whole internet, writes something with citations, runs code to check itself. - Harness: Claude Agent SDK - Model provider: Anthropic - Framework: Pydantic AI - Memory: Cognee - Retrieval: LanceDB - Web grounding: Exa - Tools & integrations: Firecrawl - Security: Guardrails AI + E2B + iron-proxy - Observability & evals: Pydantic Logfire - Runtime & deploy: Modal ```bash curl -fsSL "https://newagent.build/api/stack?harness=claude-agent-sdk&model=anthropic&framework=pydantic-ai&memory=cognee&retrieval=lancedb&search=exa&tools=firecrawl&security=guardrails-ai&security=e2b&security=iron-proxy&observability=logfire&runtime=modal&format=sh" | sh ``` ## Browser automation Drives real web apps that have no API, on behalf of a specific signed-in user. - Harness: Claude Agent SDK - Model provider: Anthropic - Framework: Mastra - Memory: Mem0 - Retrieval: pgvector - Web grounding: Firecrawl - Tools & integrations: Browserbase - Security: Descope Agentic Identity + gVisor + iron-proxy - Observability & evals: Langfuse - Runtime & deploy: Fly.io ```bash curl -fsSL "https://newagent.build/api/stack?harness=claude-agent-sdk&model=anthropic&framework=mastra&memory=mem0&retrieval=pgvector&search=firecrawl&tools=browserbase&security=descope-agentic&security=gvisor&security=iron-proxy&observability=langfuse&runtime=fly&format=sh" | sh ``` ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Guides # Defending an agent against prompt injection _Updated 2026-08-21._ If your agent reads anything written by someone who is not your user — an inbound email, a web page, a support ticket, a PDF, a repository issue — you have a prompt injection problem. There is no known complete defence, and treating detection as if it were one is the most common architectural mistake in production agents. What works is layering: reduce the attack surface, detect what is detectable, and make a successful attack worthless. ## Why detection alone is not a defence Injection detectors are classifiers, and classifiers have false negative rates. Published figures make the gap concrete: in a 2026 evaluation of agent data injection attacks, Llama Prompt Guard 2 detected 326 of 935 instruction-injection attempts — about 35% — and half of the attacks still succeeded with it in place. That is not an argument against using a detector. It is an argument against making the detector the only thing between an attacker and your data. Build on the assumption that some fraction of injections will get through, and decide now what happens when they do. In the registry: [Lakera Guard](https://newagent.build/c/lakera), [Rebuff](https://newagent.build/c/rebuff), [Llama Guard](https://newagent.build/c/llama-guard), [Invariant](https://newagent.build/c/invariant) ## Remove the payout: default-deny egress An injection is only valuable if the agent can act on it. The most common goal is exfiltration — getting your data to somewhere the attacker controls. Default-deny egress removes that: the agent can only reach destinations on an explicit allowlist, and everything else is refused. This does not prevent injection. It means a successful injection has nowhere to send anything, which in practice is most of the damage. The stronger implementations also keep credentials out of the workload entirely: the sandbox holds opaque proxy tokens, and real secrets are swapped in at the boundary, so an attacker who compromises the agent walks away with tokens that are useless anywhere else. Two details separate a real egress boundary from a checkbox. It must deny by allowlist rather than blocklist, because the destinations an attacker picks are precisely the ones nobody thought to forbid. And it must refuse a host whose resolved address falls in a denied range even when the hostname is allowed — otherwise an allowlisted name pointing at the cloud metadata endpoint reopens the whole thing. In the registry: [iron-proxy](https://newagent.build/c/iron-proxy), [Cloudflare AI Gateway](https://newagent.build/c/cf-ai-gateway) ## Contain what the agent executes If the agent writes code that then runs, the blast radius of an injection is whatever that code can do. A container is not a sandbox — shared-kernel isolation was designed to separate cooperating workloads, not to contain hostile ones. Use a kernel boundary. A user-space kernel intercepts syscalls before they reach the host and starts fast, at 10–30% overhead on I/O-heavy work. A microVM gives each sandbox its own kernel and boots in roughly 125ms with under 5 MiB of overhead, which makes per-task disposable sandboxes affordable. In the registry: [gVisor](https://newagent.build/c/gvisor), [Firecracker](https://newagent.build/c/firecracker), [Kata Containers](https://newagent.build/c/kata), [E2B](https://newagent.build/c/e2b) ## Give the agent less to steal Most damage needs privilege. An agent acting with a service account that can read every customer's data will, under injection, read every customer's data. An agent acting with a per-user token scoped to the requesting user cannot. Redact before the model sees it, not after. If personal data never enters the context window, no injection can extract it from there. In the registry: [Arcade](https://newagent.build/c/arcade), [Auth0 for AI Agents](https://newagent.build/c/auth0-ai), [Descope Agentic Identity](https://newagent.build/c/descope-agentic), [Oso](https://newagent.build/c/oso), [Microsoft Presidio](https://newagent.build/c/presidio) ## Test it like an attacker, in CI Injection defence regresses silently — a prompt change, a new tool, a model upgrade. Run a red-team suite on every change rather than auditing once at launch. Public benchmarks give you a starting corpus, and a local eval harness turns it into a test you can fail a build on. In the registry: [promptfoo](https://newagent.build/c/promptfoo) ## Common questions ### Can prompt injection be fully prevented? No. There is no known complete defence, because the model cannot reliably distinguish instructions written by your user from instructions embedded in content it was asked to read. The practical goal is to make a successful injection worthless rather than impossible: sandbox execution, deny egress by default, and scope credentials to the requesting user. ### Does egress control stop prompt injection? No, and it is worth being precise about this. Egress control does not prevent the injection from succeeding. It removes the attacker's payout — an agent that cannot reach arbitrary destinations cannot exfiltrate data, no matter what instructions it was tricked into following. ### Is a Docker container enough to sandbox an AI agent? No. Containers share the host kernel and were designed to isolate cooperating workloads, not hostile ones. If an agent executes code it wrote, use a user-space kernel such as gVisor or a microVM such as Firecracker or Kata Containers. ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # Choosing agent memory _Updated 2026-08-21._ Memory is the layer most teams add too early and evaluate least rigorously. Before choosing a product, be precise about which problem you have: session state that survives a restart, facts about a user that should persist across months, or a corpus of documents you need to search. Those are three different problems and only the second is what memory products are for. ## Start with Postgres, not a memory product A messages table and a summaries table solves more cases than most teams expect, is trivially auditable, and imposes no new vendor on your compliance review. Move to a dedicated memory product when you can name the specific query it makes fast that your own schema does not. That is not a dismissal of the category — it is the order that avoids buying an abstraction before you understand the problem it abstracts. In the registry: [Postgres (rolled by hand)](https://newagent.build/c/postgres-memory), [pgvector](https://newagent.build/c/pgvector) ## Read the benchmark numbers sceptically Agent memory has an unusually unreliable benchmark culture. Most headline scores are produced by the vendor being measured, and independent re-runs have disagreed sharply. On LoCoMo, a long-conversation question-answering benchmark, one vendor's published figure is 92.5% while an independent evaluation of the same system reports 62.47%. Another vendor's 84% claim was corrected to 58.44% in a public re-evaluation of its own paper. The lesson is not that any particular product is bad. It is that a single vendor-published number tells you almost nothing, and that the spread between evaluations is larger than the spread between products. Ask who ran the benchmark and with which judge model before you weight the result at all. In the registry: [Mem0](https://newagent.build/c/mem0), [Zep](https://newagent.build/c/zep), [LangMem](https://newagent.build/c/langmem), [Letta](https://newagent.build/c/letta) ## Retrieval and reasoning are different models of memory Most memory products embed conversation history and retrieve the nearest chunks — fast, familiar, and prone to returning things that are similar rather than things that are true. A newer approach treats memory as a reasoning problem: a model extracts preferences, beliefs and contradictions in the background, and you query it in natural language rather than searching a vector store. Pick by question shape. If you need 'what did the user say about X', retrieval is right. If you need 'what does this user actually want', a user-modelling approach fits better. In the registry: [Honcho](https://newagent.build/c/honcho), [Supermemory](https://newagent.build/c/supermemory), [Cognee](https://newagent.build/c/cognee) ## Common questions ### What is the difference between agent memory and RAG? RAG searches a corpus you own — documents, tickets, code. Memory persists what the agent learned from interacting with a particular user or over a particular task, so it still knows it tomorrow. They solve different problems and most production agents need both. ### Are agent memory benchmarks reliable? Treat them with caution. Most published LoCoMo and similar scores are produced by the vendor being measured, and independent re-runs have differed from vendor claims by as much as 30 points on the same benchmark name. Always check who ran the evaluation. ### Do I need a memory product to build an agent? No. A messages table and a summaries table in Postgres covers a large share of real cases and keeps the system auditable. Adopt a memory product when you can name the query it makes fast that your own schema cannot. ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # The self-hosted agent stack _Updated 2026-08-21._ Some requirements are not negotiable: data cannot leave the estate, no third party may see prompts, or the whole system must run in a network with no egress. Every layer of an agent has an option that satisfies this, though the trade-offs are real and worth being honest about. ## The constraint is usually narrower than it sounds Before designing for full self-hosting, establish what the requirement actually is. 'No data leaves our infrastructure' and 'no data is retained by a third party' are different constraints with very different costs. The second can often be met by a hosted provider under the right contract, or by a provider that does not retain prompts at all — which is a much cheaper path than running your own inference. Be careful with the distinction: a provider that does not log your prompts is still a third party on the wire. That satisfies a retention requirement, not a residency one. In the registry: [Venice AI](https://newagent.build/c/venice), [Amazon Bedrock](https://newagent.build/c/bedrock), [Google Vertex AI](https://newagent.build/c/vertex) ## Inference on your own hardware For genuine on-premise inference the choice is between simplicity and throughput. A local runner is the fastest path to something working on one machine. A dedicated serving stack is what you want once concurrency matters, and it is where the open-weight models become genuinely practical. Open weights are also a hedge worth taking even when you use a hosted API: if the hosted option later becomes unacceptable, the same model runs on your own GPUs without a rewrite. In the registry: [Ollama](https://newagent.build/c/ollama), [vLLM](https://newagent.build/c/vllm), [LiteLLM](https://newagent.build/c/litellm), [DeepSeek](https://newagent.build/c/deepseek) ## Everything else has an open-source answer Memory and retrieval collapse into one dependency you probably already run. Web grounding, the layer teams assume forces a third party, has a self-hosted metasearch option. Observability, guardrails and containment are all available as software you deploy yourself. In the registry: [pgvector](https://newagent.build/c/pgvector), [SearXNG](https://newagent.build/c/searxng), [Langfuse](https://newagent.build/c/langfuse), [Microsoft Presidio](https://newagent.build/c/presidio), [NeMo Guardrails](https://newagent.build/c/nemo-guardrails), [Kata Containers](https://newagent.build/c/kata), [iron-proxy](https://newagent.build/c/iron-proxy) ## What self-hosting costs you Being honest about the trade: you take on GPU capacity planning, model upgrades, and the operational burden of every component in the stack. Open-weight models trail the frontier on the hardest reasoning tasks, though the gap on coding benchmarks has narrowed to roughly a point at the top of the leaderboards. Budget for the operational load, not just the licence savings. ## Common questions ### Can an AI agent run entirely on-premise? Yes. Every layer has a self-hostable option: vLLM or Ollama for inference, Postgres with pgvector for memory and retrieval, SearXNG for web grounding, Langfuse for observability, Presidio and NeMo Guardrails for guardrails, and Kata Containers with an egress proxy for containment. ### Is a no-logging hosted provider enough for compliance? It depends which requirement you have. A provider that does not retain prompts satisfies a data-retention requirement but not a data-residency one — the data still crosses a third-party network. If your obligation is residency or air-gapping, you need inference on your own hardware. ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. --- # What is an agent harness? _Updated 2026-08-21._ The harness is the layer most agent directories omit entirely, which is odd, because in practice it is the first thing you choose. The model thinks; the harness is what reads files, runs commands, calls tools, keeps the session, and decides when to stop and ask a human. ## Harness versus framework A framework is a library you write orchestration in — you own the loop, and the library gives you graph execution, checkpointing or type-safe outputs. A harness is a loop that already exists, which you configure and drive. The practical test: if you are writing the while-loop that calls the model and dispatches tools, you are using a framework. If something else runs that loop and you supply tools, prompts and permissions, you are using a harness. In the registry: [Claude Code](https://newagent.build/c/claude-code), [Claude Agent SDK](https://newagent.build/c/claude-agent-sdk), [LangGraph](https://newagent.build/c/langgraph), [Mastra](https://newagent.build/c/mastra) ## What the harness decides for you Choosing a harness settles more than it appears: how tools are declared, how sessions persist, what the approval model looks like, how context is compacted when it overflows, and how far the agent can go before a human intervenes. These are the properties you will care most about in production, and they are harder to change later than the model. ## The 2026 harness wave The category expanded sharply in 2026. Alongside the vendor coding agents, several harnesses now treat the loop itself as the product: one where prompts, skills, sub-agents and memory are state the agent can create and revise mid-run; one where every part including the agent loop is a swappable plugin; and several self-hosted gateways that put a single agent into Slack, Telegram and WhatsApp at once. Two properties are worth weighing directly. Whether the harness is model-agnostic, since a harness tied to one vendor's models is a harness tied to that vendor's pricing. And whether it improves with use — some now write their own skills from experience and carry them into later sessions. In the registry: [Hermes Agent](https://newagent.build/c/hermes-agent), [Prime Agent](https://newagent.build/c/prime-agent), [DeepSeek Harness](https://newagent.build/c/deepseek-harness), [OpenClaw](https://newagent.build/c/openclaw), [OpenCode](https://newagent.build/c/opencode) ## Common questions ### What is an agent harness? An agent harness is the runtime that executes the agent loop: it reads files, runs shell commands, calls tools, manages the session and handles approvals. The model supplies the reasoning; the harness connects it to the real world. ### What is the difference between an agent harness and an agent framework? A framework is a library you use to build your own loop. A harness is a loop that already exists, which you configure. If you are writing the code that calls the model and dispatches tools, that is a framework; if something else runs that loop for you, that is a harness. ### Do I need a harness to build an AI agent? No, but it is usually the faster path. Writing your own loop with a model API and a tool dispatcher is entirely viable and gives you full control. A harness saves you the session management, context compaction and approval plumbing you would otherwise build yourself. ## Machine interfaces - `GET https://newagent.build/api/registry` — every component as JSON - `GET https://newagent.build/api/vendors` — vendors, with the repo and releases feed to watch - `GET https://newagent.build/api/advise?q=` — recommended stack for a description - `GET https://newagent.build/api/stack?=&format=sh|json|md|yml|agents|env` — a stack as files - `GET https://newagent.build/llms.txt` — the whole registry in one fetch Any page here also returns markdown if you send `Accept: text/markdown`. ---