Composio powers 1000+ toolkits, tool search, context management, authentication, and a sandboxed workbench to help you build AI agents that turn intent into action.
Mirrored from https://github.com/Camo67/composio.git
Synced Nov 21, 07:15 UTC
docs: clarify managed auth for custom tools and improve discovery (P0) (#5154)
## Why Addresses the P0 documentation recommendations for [DEVREL-229](https://linear.app/composio/issue/DEVREL-229/clarify-composios-auth-only-positioning-for-multi-tenant-agents-in). Agents conflate Composio's prebuilt tools, shared OAuth apps, and credential custody, and struggle to confirm managed authentication with application-owned tools. ## Changes - Add a complete Python, TypeScript, and REST walkthrough for connecting a tenant/member's GitHub account and making an account-specific authenticated proxy request from application-owned code. - Explain connection reuse, callback ownership checks, expired/revoked connections, provider-domain restrictions, and the separate proxy permission. - State the supported boundary explicitly: no predefined tools, MCP, or LLM required; Composio still executes the authenticated provider request; account reads are not a token-export workflow. - Make managed-versus-custom OAuth app guidance requirements-based for production, including the canonical and curated production KB sources and regenerated presentation. - Link the guide from authentication navigation, the auth overview, managed-app catalog, B2B guide, proxy/custom-tools pages, and `llms.txt`. - Scope the agent Markdown appendix's Native Tools/MCP wording to agent tool delivery rather than treating it as an exhaustive list of execution paths. ## Evidence Existing hosted Gauge runs reviewed through the CLI: - `cmvtk8d6p000j01fy2dulqqrt`: Codex discovered the authentication/token-custody docs and accurately described Proxy Execute and token redaction. - `cmvtk8d6p000i01fyi15t1nnw`: Claude explicitly did no research and repeated the stale token-retrieval claim. - `cmvh6unx3000l01jqpnhiyfpf`: the managed-auth catalog did not resolve the agent's framework-decoupling question. These observations motivate clearer docs and discovery, not a claim of measured improvement. No new Gauge run or before/after comparison is claimed. ## Verification - `bun run check:code-examples`: passed, 446 files. - `bun run check:api-snippets`: passed, 446 files. - `bun run check:python-snippets`: passed, 350 snippets. - Type-checked the new guide's TypeScript fence with the existing snippet checker's Twoslash API against the installed docs SDK: one snippet, zero diagnostics. - `bun run lint:links`: passed, including SDK links. - `bun run test --timeout 30000`: 814 passed, zero failed. - `bun run lint`: passed with 63 warnings in unchanged files. - Browser smoke: rendered the new guide, switched Python/TypeScript/API tabs, checked sidebar discovery, and fetched its Markdown, `llms.txt`, auth overview, managed-auth catalog, and production KB. Initial verification exposed missing product-sidebar registration and an incidental test that pinned the old first auth page. Registered the guide, removed that positional snapshot rather than repinning it, and kept the navigation coverage invariants. Kept `llms.txt` within its existing length bound. Full `GITHUB_ACTIONS=true bun run build` did not finish its optimized-build phase before the local 600-second deadline. Development rendering and the targeted TypeScript check passed; no successful full production build is claimed. Authenticated OAuth/provider execution was not performed; it requires credentials and user consent. ## Scope Docs only; no SDK/API behavior or package releases. Companion P1 PR: #5153 (B2B execution-choice guidance, identity-boundary wording, and a sourced Nango comparison). Both target `next` independently. Review the full diff before merging.
No changes.