Composio

Composio translates tool definitions into framework-specific formats (LangChain tool_choice, CrewAI @tool decorators, AutoGen function_map, OpenAI function_calling) via provider packages that wrap the core SDK. Each provider package implements a framework adapter that converts Composio's OpenAPI-based tool schemas into native function-calling conventions, enabling agents to discover and invoke tools without framework-specific boilerplate. The routing happens through a session-based tool router that maintains authentication context across framework calls.

Composio's authentication system handles OAuth 2.0 flows, API key storage, and custom auth schemes through a centralized credential manager at the backend API. When an agent needs to call a tool (e.g., GitHub API), Composio retrieves the stored credential from the backend, automatically refreshes OAuth tokens if expired, and injects the auth header into the outgoing request. Credentials are stored server-side with encryption, and the SDK never handles raw secrets locally—only credential IDs are passed to agents.

Composio provides a CLI (@composio/cli for TypeScript, composio CLI for Python) that enables developers to explore toolkits, test tool execution locally, and manage authentication without writing code. The CLI includes commands to list available toolkits, view tool schemas, test tool calls with sample parameters, and authenticate with external services. The CLI is built as a binary (via pkg for Node.js, PyInstaller for Python) and can be distributed standalone without requiring SDK installation.

Composio manages toolkit versions independently—each toolkit (GitHub, Slack, Jira, etc.) has its own version number and release cycle. Agents can pin specific toolkit versions, enabling controlled updates without forcing all toolkits to upgrade together. The backend API supports multiple toolkit versions simultaneously, allowing gradual migration from old to new schemas. Breaking changes in toolkit schemas trigger major version bumps, and the SDK provides deprecation warnings for outdated versions.

Composio provides framework-specific provider packages (composio-langchain, composio-crewai, @composio/langchain, etc.) that implement native integration patterns for each framework. For LangChain, the provider exports StructuredTool objects that integrate with LangChain's tool_choice mechanism. For CrewAI, the provider exports decorated functions that work with CrewAI's @tool decorator. For AutoGen, the provider exports function_map dictionaries. Each provider package handles framework-specific details (tool calling conventions, error handling, async patterns) transparently.

Composio uses sessions to maintain authentication state and tool availability across multiple agent calls. When an agent creates a session, Composio binds a set of connected accounts (authenticated credentials) to that session. The session-based tool router then ensures that all tool invocations within that session use the correct credentials. Sessions can be scoped to users, conversations, or workflows, enabling multi-tenant isolation and per-user tool access control without re-authenticating on each call.

Composio maintains a registry of 500+ pre-built toolkits, each defined as OpenAPI schemas. When an agent requests tools from a toolkit (e.g., GitHub), Composio serves the OpenAPI schema, which includes operation descriptions, parameter types, and response schemas. The SDK automatically converts these schemas into agent-readable documentation (function descriptions, parameter hints) and generates tool discovery endpoints that agents can query to find available actions. Toolkit versions are managed independently, allowing agents to pin specific versions without affecting other toolkits.

Composio's trigger engine enables agents to subscribe to real-time events from external services (e.g., 'new GitHub issue', 'Slack message in channel') via webhooks and WebSocket connections (Pusher). When an event occurs, Composio's backend receives the webhook, matches it to subscribed agents, and delivers the event payload to the agent's execution context. Agents can define trigger handlers that automatically invoke tool actions in response to events, enabling reactive workflows without polling.

Composio's file service abstracts file I/O for tools that require file uploads or downloads (e.g., uploading a document to Google Drive, downloading a GitHub repository). When an agent needs to upload a file, Composio generates a presigned S3 URL, the agent uploads the file directly to S3, and Composio passes the S3 reference to the target service API. For downloads, Composio retrieves files from the service, stores them in S3, and provides presigned download URLs to the agent. This approach avoids loading entire files into agent memory and reduces latency for large file operations.

Composio allows developers to define custom tools beyond the 500+ pre-built integrations using the custom tool API. Custom tools are defined via OpenAPI schemas or Python/TypeScript function signatures, and can be added to a toolkit. The modifier system enables post-processing of tool execution—modifiers can transform tool inputs before execution, intercept outputs, add logging, or implement custom error handling. Modifiers are composable and can be chained together, enabling flexible tool behavior customization without forking the SDK.

Composio provides an MCP server at mcp.composio.dev that exposes all 500+ toolkits via the Model Context Protocol standard. Agents or LLM clients that support MCP can connect to this server and discover tools without using Composio's SDK directly. The MCP server handles authentication, tool routing, and execution, translating MCP requests into Composio's internal tool execution pipeline. This enables integration with MCP-compatible clients (Claude Desktop, custom MCP clients) without SDK dependency.

Composio's execution engine implements rate limiting and error handling transparently. When a tool call hits a rate limit (e.g., GitHub API 60 requests/hour), Composio automatically retries with exponential backoff. For transient errors (network timeouts, 5xx responses), Composio retries up to a configurable limit. For permanent errors (4xx responses, invalid parameters), Composio returns the error immediately without retry. Rate limit state is tracked per-session and per-service, enabling fair resource allocation across concurrent agents.

Composio provides parallel Python and TypeScript SDKs with identical core functionality but language-specific implementations. The Python SDK uses async/await patterns and integrates with CrewAI, LangChain, and Anthropic's Python clients. The TypeScript SDK uses Promise-based async and integrates with LangChain.js, Vercel AI SDK, and Node.js-based frameworks. Each SDK has provider packages (e.g., @composio/langchain, composio-langchain) that implement framework adapters. This dual-SDK approach enables language-agnostic tool integration without forcing teams to adopt a single language.