Skip to main content

MCP & Connecting to Tools

Advanced

The Model Context Protocol (MCP) is the open standard for connecting AI to external tools and data. On the API, MCP lets you give Claude access to a whole server's worth of tools by reference — instead of defining each tool by hand.

MCP vs hand-defined tools

Tool use (custom)MCP
You defineEach tool's schema, and run it yourselfA connection to a server that publishes tools
Best forA few bespoke functions in your appReusing existing integrations (GitHub, DBs, browsers, SaaS)
AuthYour codeOAuth or native keys to the server

They coexist: define your app-specific tools directly, and pull in ready-made capabilities via MCP.

Declaring a server

You attach MCP server(s) to a request/agent (URL or command, plus auth). Claude can then call the server's tools within the agent loop, same as any tool. The exact request field and auth flow are in the official docs.

Same standard, three surfaces

  • API (this page) — MCP servers on your requests/agents.
  • Claude Code — MCP servers in your dev sessions.
  • The apps — MCP powers Connectors.

Learn it once; it transfers.

Trust

:::warning An MCP server is code + access Only connect servers you trust, scope them to least privilege, and remember external content they return can carry prompt injection. Review third-party servers first — Reviewing Third-Party Code. :::

Next