// the find
ComposioHQ/composio
Composio powers 1000+ toolkits, tool search, context management, authentication, and a sandboxed workbench to help you build AI agents that turn intent into action.
Composio is a hosted platform that gives AI agents pre-authenticated access to 1000+ third-party apps (Gmail, Slack, GitHub, etc.) without you having to implement OAuth flows yourself. You install an SDK, create a per-user session, and get back tool definitions your agent can call. The session abstraction handles auth, rate limits, and tool discovery at runtime so you don't dump 1000 tool schemas into context at once.
The session-scoped meta-tool approach is genuinely smart: instead of loading every available tool upfront, the agent discovers and authenticates tools on demand, which keeps context size manageable. Provider adapters cover nearly every mainstream framework (OpenAI Agents, Claude Agent SDK, LangChain, Vercel AI, Mastra, LlamaIndex), so you're not locked into one ecosystem. The repo ships both TypeScript and Python SDKs with parity, tested against real framework versions via nightly CI, which is more disciplined than most multi-language OSS projects. The MCP endpoint per session is a practical escape hatch — any MCP client works without the SDK.
The core value (pre-authenticated toolkits) is a hosted service, not something you run yourself — the SDK is a client to Composio's cloud, so you're betting on their uptime, pricing, and data-handling decisions for every tool call your agent makes. With 1000+ toolkits, the quality variance is enormous and there's no obvious signal for which integrations are production-grade vs. barely-working stubs. The Python SDK versioning is separate from the TypeScript SDK (different changelogs, different release cadences), so cross-SDK parity is a stated goal that will inevitably lag. Error handling in agentic loops is opaque — when a tool call silently fails or returns a wrong result, debugging through Composio's abstraction layer adds a layer you don't own.