GitHub ↗

What is an MCP server?


Short answer: an MCP server is a small program that hands an AI assistant one specific capability — reading a database, driving a browser, calling an API — through a standard interface. It contains no model and no intelligence of its own. It advertises a list of tools, and when the assistant decides to use one, the server does the work and returns the result. Most run locally on your own machine.

(If you want the protocol itself rather than the component, start with Model Context Protocol, explained.)

What a server actually is

Strip away the terminology and a typical MCP server is: a process that starts when your AI client starts, says "here are the things I can do," waits, and executes requests. That's it. It's closer to a plugin than to a web server, despite the name — "server" here means "serves capabilities," not "runs in a data centre."

Two things follow from that, and they're the ones worth internalising:

  1. It runs with your permissions. A local server can do anything you could do from a terminal on that machine. That's the source of both its power and its risk.
  2. It has no model. The intelligence stays in the client. The server is the hands, not the brain.

What runs where

Local (stdio)Remote (HTTP)
Runs onYour machineSomeone else's
Can reachYour files, your network, your logged-in sessionsOnly what you send it
Typical useFilesystem, git, databases, your browserHosted SaaS APIs
Data exposureStays localLeaves your machine

If a capability needs your environment — your files, your VPN, the browser you're signed into — it has to be local. That's not a preference, it's a constraint.

Adding one

For most clients it's a config entry naming a command to run. In Claude Code, for example, adding a server is a single claude mcp add command; Cursor and VS Code use a small JSON block. You don't build anything — you point your client at a program and restart it. Servers are increasingly discoverable through the official MCP registry rather than word of mouth.

How to tell a good server from a bad one

Since a local server runs with your permissions, this matters more than feature lists:

An example: the browser

Browser MCP is a local stdio server plus a Chrome extension. Its single job is to let an agent operate the real, already-logged-in Chrome you use — 34 tools for navigating, reading, clicking, filling and screenshotting. It has to be local, because the whole point is your own browser session, and nothing it reads leaves your machine. It's a fair example of the pattern: narrow job, local by necessity, source you can read.

FAQ

Is an MCP server the same as an API? No. An API is a service you call; an MCP server is an adapter that makes some capability (which might be an API, or your filesystem, or your browser) available to an AI client in a standard shape.

Do MCP servers cost money? The protocol and most servers are free and open source. Some wrap paid services, in which case you pay that service.

How many can I run at once? Several — clients let you register multiple servers, and the assistant picks tools across them.

Can an MCP server see my whole computer? A local one runs with your user permissions, so treat installing it like installing any local software: prefer readable source and a narrow scope.

Which should I install first? Whichever removes a chore you actually repeat. If that's anything involving a logged-in website, start here.