MCP server
An MCP server is a program that exposes capabilities to AI applications over the Model Context Protocol through three building blocks: tools (schema-defined functions the model can call), resources (read-only data such as documents or schemas), and prompts (reusable instruction templates). For an API producer, an MCP server is the protocol-native package that lets any MCP-capable agent call the API at runtime.
Mechanically, a server speaks JSON-RPC 2.0 with an MCP client, declares its tools with JSON Schema-typed inputs, and can run locally alongside the host application or remotely over HTTP. The model decides when to invoke a tool; the host executes the call and returns the result into the model's context. Common examples range from file-system and database servers to servers wrapping a single vendor's API.
Shipping one changes the runtime story: an agent no longer has to construct raw HTTP requests from prose documentation, because the server presents typed operations. It changes less than producers hope about being chosen, since agents and developers still have to discover that the server exists — via a manifest at a well-known path or a registry listing — and still have to understand what its tools do from the descriptions the producer wrote. A server with vague tool descriptions reproduces the documentation problem one layer down.
Discry registers MCP servers through the discovery dimension: the .well-known/mcp.json manifest check and the MCP registry listing check both credit an API whose MCP surface is publicly findable. The behavior of the server's tools is not yet part of the measurement; the methodology discloses that boundary explicitly.
What does an AI agent make of your API?
Find out in about a minute — no signup.
Discry your API — free