MCP & Connecting to Tools
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 define | Each tool's schema, and run it yourself | A connection to a server that publishes tools |
| Best for | A few bespoke functions in your app | Reusing existing integrations (GitHub, DBs, browsers, SaaS) |
| Auth | Your code | OAuth 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. :::