What is an MCP server?

An MCP server is software that exposes tools, resources, or prompts through the Model Context Protocol so an AI client can interact with an external system. A server may run locally over stdio or remotely over HTTP. The client discovers the server's capabilities through MCP and calls them through the protocol.

Every server can expose any combination of tools, resources, and prompts. A database server might expose a query tool; a search server might expose a search tool and a results resource. Nothing in the protocol requires a server to expose all three.

A local server runs on the same machine as the client and communicates over stdio: the client spawns the server process and exchanges JSON-RPC over standard input and output. This is the simplest deployment and requires no network exposure or authentication.

A remote server runs on another host and is reachable over HTTP. A client establishes a session over the network. Authentication is the server's responsibility; the specification supports OAuth-based authorization for protected resources.

A client learns a server's capabilities by sending tools/list, resources/list, and prompts/list and reading the returned schemas. There is no global registry requirement; discovery comes from the protocol conversation itself.

Before using a server, verify what tools or resources it actually exposes, whether it needs credentials, whether it can mutate state, and whether it follows a transport and authentication you can support. Read the server's source or documentation rather than trusting a name alone.