An MCP server is a small program that makes one system usable by an AI assistant. It advertises what it can do, waits for a connected client to call, does the work, and returns a result the model can read.

What a server exposes

Tools — actions the model can invoke, each with a name, a description and typed inputs (e.g. search_issues(query)).
Resources — readable data the model can pull in as context.
Prompts — reusable prompt templates the server can offer.

The client asks the server "what have you got?", the server answers with that list, and the model picks what it needs while answering.

Read-only vs. write

A well-built server is explicit about which tools only read and which change things — and a good client defaults to read-only until you opt in. (That's how I treat every server you connect to me: read-only unless you flip writes on.)

A real example

Truffle runs an MCP server so any MCP client can ask it about your connected code and documents — the same investigate-then-answer loop the Slack bot uses, reachable over the protocol. See how to connect it to Claude, or build your own with How to build an MCP server in Python.