An MCP server is a small program that exposes tools and data to an AI model through a standard protocol. Think of it as a universal adapter: instead of every AI app building a custom integration for every service, the service ships one MCP server and any MCP-capable client can use it.
The one-sentence version
An MCP server lets an AI do things and read things outside its own text box — open a file, query a database, search the web, create a GitHub issue — by offering those abilities as callable "tools."
Client vs server
MCP has two sides:
- MCP client — the AI app you use: Claude Desktop, Cursor, Windsurf, VS Code. It hosts the model and connects to servers.
- MCP server — the connector for one capability or system: a filesystem server, a Postgres server, a GitHub server.
One client can connect to many servers at once. That is the whole point — compose abilities like Lego bricks.
What a server actually exposes
- Tools — actions the model can call (
run_query,create_issue,read_file). - Resources — data the model can read (files, records).
- Prompts — reusable prompt templates the server provides.
When you ask Claude "what's in my projects folder," the client asks the filesystem server to list files, gets the result, and the model answers from it.
How it connects
Servers run two main ways:
- stdio — a local process the client launches (most common; configured with
commandandargs). - HTTP/SSE — a remote server the client connects to over the network (good for hosted services).
See MCP config file location for where clients store this.
Why it took off
Before MCP, every AI app reinvented integrations. The Model Context Protocol standardised the interface, so the ecosystem exploded — from a few dozen servers in early 2025 to thousands by 2026. For the protocol itself, see what is MCP, and how it compares to alternatives in MCP vs API.
Try one
The easiest first server is the filesystem one — see the filesystem guide — or browse ready-made loadouts. Want to build your own? Start with build your first MCP server.