Turn your backend into secure tools for AI agents.

Add an mcp block and any interface becomes a Model Context Protocol tool that Claude, Cursor and other agents can call — with the same authentication and validation as a normal HTTP route.

One config, an API and an agent tool

The interface you already built to serve HTTP can be exposed as an MCP tool at the same time. No second implementation, no separate tool server — the engine handles the MCP protocol for you.

• Works with Claude Desktop, Claude Code, Cursor and any MCP client
• Reuses your per-request JWT / API-key auth
• Same assertions and transforms guard agent calls

Name the server, not just the tools

Declare your server's title and description alongside the tools, and every client and registry reads them off the MCP handshake. Declare more than one and a single deployment publishes several endpoints, each with its own identity and its own subset of tools.

• Listing title and description come from the same config as the tools
• One identity covers tools spread across many configs
• Split customer-facing and internal tools onto separate MCP endpoints

Safe by construction

Agents only get what you declare. Your assertions, auth and network policy apply to MCP calls exactly as they do to HTTP — so an agent can't reach anything a normal client couldn't.

Frequently asked questions

• What is MCP? The Model Context Protocol is an open standard for giving AI clients callable tools. An MCP tool is something an agent can invoke — reading a row, kicking off a job — rather than text it has to guess from.
• How does auth work for MCP tools? Exactly as it does for HTTP. An MCP-exposed interface reuses the same per-request JWT or API-key checks, assertions and network policy, so an agent can't reach anything a normal client couldn't.
• Which MCP clients are supported? Any MCP client — Claude Desktop, Claude Code and Cursor among them. Air Pipe implements the protocol, so the client is your choice.
• Can I control how my server appears in a client or registry? Yes. Your config declares the server's title and instructions, and Air Pipe returns them on the MCP handshake — the only place registries such as mcp.so, Glama, Smithery and PulseMCP can read a remote server's listing description from. You can declare several named servers too, so one deployment publishes more than one endpoint, each with its own identity and tools.
• Is this the same as managing Air Pipe from an AI client? No, and it's worth separating them. This page is about exposing your own interfaces as tools an agent can call. Air Pipe also has a management MCP for driving the platform itself from an AI client — a different thing in the other direction.