How to install an MCP server on your own computer

What you need first

Three things, just once.

Node.js installed. It's the engine that runs most MCP servers. It's downloaded for free from the official website; during setup you'll use the npx command, which comes with Node.
An AI that supports MCP. For example Claude in its desktop app or a compatible code editor. If the app doesn't have a section for MCP servers, you can't connect any.
Knowing what you want to connect. To read your files you need the filesystem server; for other purposes, the right server. Start from the problem.

How to do it

The principle is the same on Windows, Mac and Linux; only where the configuration file is located changes.

Install Node.js from the official website, if you don't have it. To check whether it's already there, open the terminal and type node --version: if it replies with a number, you're set.
Open your assistant's configuration file. In desktop apps it's usually reached from the settings, under developer or integrations, which opens the right file.
Add the server. The operative syntax (example with the filesystem server giving access to a folder):

{
  "mcpServers": {
    "documents": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/yourname/Documents"]
    }
  }
}

The npx -y part tells the computer to download and start the server on the fly, without you installing it by hand. The last value is the folder the server will be able to see: put a precise one.

Restart the assistant completely (closing the window isn't enough). At restart it reads the configuration and connects to the server.
Verify it's there. Ask the AI to list the tools it has available, or look for the MCP tools icon in the app. If it appears, the connection succeeded.

A concrete example

Luca wants the AI to read his notes folder. He checks Node with node --version: it's not there, he installs it from the official website in two minutes. He opens the app's configuration from the developer settings, pastes the filesystem server block and changes the path to that of his notes folder. He saves, closes and reopens the app.

At restart he asks: "what tools do you have now?". The assistant lists the filesystem functions. Luca writes "summarize yesterday's meeting notes" and the AI reads the right file and replies. Total time: ten minutes, of which eight to install Node the first time. From the second server installation, he'll skip that step.

When it does NOT work (and how to fix it)

If the server doesn't appear after the restart

Almost always it's the configuration file written wrong. An assistant won't load a server if the JSON (the file's format, made of brackets and commas) isn't valid: an extra comma or a missing bracket is enough. Paste the file's contents into the AI itself and ask "is this JSON valid? fix it": it's the fastest way to find the error.

If you get the "command not found" error or similar

It means Node isn't installed or the computer can't find it. Reinstall Node.js from the official website and, importantly, close and reopen the terminal and the app after installation, so they recognize the new command.

If the server starts but doesn't see the files

Check the folder path: it must be the exact one and written correctly (on Windows the slashes must be handled carefully). If you have doubts about the path, copy it directly from the folder's address bar instead of typing it by hand.

A tip from someone who really uses it

Before connecting a server to real data, test it on a test folder with a few dummy files. See how the AI behaves when it "sees" the files, get comfortable, and only then point the server to the real folder. Getting the configuration wrong on dummy files costs nothing; getting it wrong on work documents can.

Frequently asked questions

Do I have to install a server for everything I want to do?

Yes, one server per type of capability: one for files, one for the web, and so on. Each is added with the same method, an entry in the configuration file. It's best to start with one and add the others when they're really needed.

Can I test a server before connecting it to the AI?

Yes. There's an official tool called MCP Inspector that opens a server in a browser window and shows you what it can do, without connecting it to the assistant. It's useful for understanding a new server before giving it access to your data.

Do I have to know how to program to install an MCP server?

Here lies the misunderstanding to clear up: you don't write code, you paste a configuration and change a path. It's closer to changing an app's settings than to programming. The technical part has already been done by whoever wrote the server; what's left for you is telling it where to look and restarting the assistant.

Quick answer