Composio MCP

Composio MCP is a managed gateway that connects AI assistants and agents to real tools and accounts. One MCP endpoint exposes more than 1,500 apps with authentication, permissions, triggers, and execution handled by Composio. The same platform offers SDKs for building agents, a command line tool for coding assistants, and a sandbox where code can run on large jobs. Teams can share accounts, skills, and budgets, and administrators get visibility into what each agent is allowed to do.

Key Features

• A single managed MCP endpoint covering more than 1,500 tools, with OAuth consent, token storage, refresh, and scopes handled for every connected account.
• Works with Claude, ChatGPT, Codex, Cursor, Grok, Hermes, OpenClaw, and any MCP client.
• SDKs for OpenAI, Anthropic, Vercel AI SDK, Google, and LangChain.
• Triggers that deliver events from connected apps to your webhook, plus sessions for running tools in context.
• A sandbox where the agent runs code in a private workspace and returns only the answer.
• Firewall protection that blocks instructions hidden in tool results before the agent reads them.
• Per-action permissions, audit records, team wallets, and enterprise controls such as SAML single sign-on, SCIM, customer-managed keys, and self-hosted deployment.

Who is it for

• Developers building agents that need production-ready tool integrations and auth.
• Teams connecting Claude, ChatGPT, or Codex to company accounts such as Gmail, Slack, and Notion.
• Platform and IT teams that need one place to govern every agent in the company.
• Startups that want to ship integrations without maintaining OAuth flows for each app.
• Enterprises evaluating deployment inside their own VPC or fully self-hosted.

What can you do with Composio MCP

• Add an MCP endpoint to an AI client and call tools without building each integration.
• Connect user accounts so an assistant can read, send, or update data within granted permissions.
• Run code in a sandbox when a task involves thousands of rows or heavy processing.
• Set triggers that react to events in connected apps and call your webhook.
• Share apps, skills, and a budget across a team from one setup.
• Review every tool call, the account used, and whether an action was allowed or blocked.

How does Composio MCP work

1. Create a free account and connect the accounts your assistant should use.
2. Add the Composio MCP endpoint to Claude, ChatGPT, Cursor, or another MCP client.
3. Choose which tools and actions the assistant can reach, with permissions set per action.
4. Let the agent call tools; Composio handles authentication, execution, and logging.
5. Review activity in the dashboard, adjust policies, and add team members or budgets as needed.

Pricing

• Hobby: $0, with 100,000 tool calls and 50,000 triggers per month, unlimited connected accounts, and 3 team members.
• Pro: $29 per month, with everything in Hobby, a $29 monthly usage credit, unlimited team members, spend controls, and all add-ons.
• Enterprise: custom pricing with committed volume and discounts, annual invoicing, SAML single sign-on and SCIM, customer-managed keys, white-labeling, a data processing agreement, a business associate agreement, and dedicated support with a service-level agreement.
• Usage beyond the included amounts is billed on the rate card, and Pro accounts can set spend caps to avoid surprises.

FAQ

Is the free plan really free?

Yes. The Hobby plan is hard-capped, needs no credit card, and never produces a surprise bill; when a cap is reached, usage pauses until the next month or an upgrade.

Which AI clients work with Composio?

It works with Claude, ChatGPT, Codex, Cursor, Grok, Hermes, OpenClaw, and any MCP client, and it also provides SDKs for several agent frameworks.

How does authentication work?

Composio manages OAuth consent, token storage, refresh, and scopes for each connected account, so the model never sees a password and users control what each app can do.

What is the Firewall feature?

Firewall checks tool requests and results against your policies and blocks instructions hidden in tool results before the agent reads them; on Pro it is billed per million tokens evaluated.

Can I deploy it in my own infrastructure?

Enterprise plans support deployment in a virtual private cloud or fully self-hosted environments, with zero data retention options and customer-managed keys.