An MCP server is a small program that exposes tools, data and actions to an AI agent over the Model Context Protocol, so the agent can do things like query a database, browse the web or open a pull request.
An AI model on its own can only produce text. An MCP server is the bridge that lets it actually do something: it advertises a set of tools (functions with typed inputs and outputs) that a compatible client like Claude Code or Cursor can call on the model's behalf. Ask the agent to "check the failing test and open a PR" and, with the right servers connected, it can read the repo, run the tests and push a branch.
Each server usually wraps one system: a GitHub server exposes repos, issues and pull requests; a Postgres server exposes queries and schema; a browser server exposes navigation and clicks. Because they all speak the same protocol, you can mix and match them, and any MCP-aware client can use any server without custom glue code.
Most servers run locally over stdio or as a small HTTP service, and install with a single command. Connect the ones your workflow needs and the agent gains those capabilities for the session.
Most publish an npm or PyPI package and start with one command; the client connects to it over stdio or HTTP. Each tool page here shows the exact install command.
Anything its author wraps, query databases, control a browser, call an API, read files, manage GitHub. The agent calls the server's tools as needed.
Claude Code, Cursor, and a growing list of agent tools. Because MCP is an open protocol, one server works across all of them.