MCP servers: how an AI gets hands for your other apps

Skip to the how-to Start with the idea

First, the words

Every word on this page that might trip you up is explained here. Come back whenever you need to; the glossary has more.

AI app

The program you chat with, such as Claude, ChatGPT, or an AI helper inside a code editor.

MCP

Model Context Protocol. An agreed way for AI apps and other tools to talk. "Protocol" just means "a set of rules both sides follow," like the rules for addressing a letter.

MCP server

A small program that gives the AI access to one thing, such as your calendar or a folder, and tells it what it can do there.

Host

The AI app that uses MCP servers. Claude Desktop, Claude Code and Visual Studio Code are hosts.

Client

The little connector inside the host that keeps the line open to one server. One per server. You never see it.

Tool

An action a server offers, such as "search flights" or "create calendar event".

Local / remote

Local: the server runs on your own computer. Remote: it runs on the internet, on the service's machines.

API

Application programming interface. A doorway that lets one program talk to another. Many MCP servers are a friendly wrapper around a service's API.

Connector

Claude's name for a remote MCP server you've added in its settings.

The idea in one picture

On its own, an AI only knows what it learned in training and what you paste into the chat. It can’t look at today’s calendar or open a file on your computer. An MCP server is the bridge.

The official MCP site puts it this way: think of MCP like a USB-C port for AI apps. Before USB-C, every device had its own charger. Before MCP, every AI app needed its own custom-built connection to every tool. Now each side learns one plug.

So who builds the servers? Usually the company behind the tool (Notion, Sentry and GitHub offer their own), or the open-source community. You mostly just plug them in.

Who’s who: host, client, server

There are only four players, and you are one of them. Follow the arrows from left to right (top to bottom on a phone):

1. You

   Type a request in plain words.

2. The AI app (the host)

   Claude, ChatGPT, VS Code…

   • client
   • client
   • client

   One client per server, inside the app.

3. MCP servers

   • Calendar server
   • Files server
   • GitHub server

   Each knows one thing well.

4. The real thing

   • Your calendar
   • Folders on your PC
   • GitHub.com

An everyday comparison: the host is a restaurant, the clients are its waiters and each server is a kitchen station (grill, salads, desserts). You only talk to your waiter. The waiter knows which station to ask, brings back the dish, and you never have to walk into the kitchen.

What a server can offer: three building blocks

Every MCP server offers some mix of three things. The difference between them is simply who decides to use them.

Most servers are mostly tools. If you remember one block, remember that one: tools are how the AI gets hands.

What happens when you ask: one request, step by step

Here is a real-looking example. You’ve added a calendar server to your AI app, and you ask about Friday. Follow it from the top.

1. 1

   You type: "Am I free Friday afternoon? If so, block 2 to 4 pm for the dentist."

2. 2

   The app has already asked each server "what can you do?" The calendar server answered with its tools, such as list_events and create_event.

3. 3

   The AI picks list_events for Friday. The server reads your calendar and sends back the list.

4. 4

   Friday 2 to 4 pm is empty, so the AI wants to call create_event.

5. 5

   The app asks you: "Allow create_event?" Do you approve?

   This is your safety switch. Read what it's about to do.

   No

   Nothing is changed. The AI tells you it stopped.

6. 6

   The server adds the event to your real calendar and reports back: done.

7. 7

   The AI answers: "You were free. I've added 'Dentist', Friday 2 to 4 pm."

The AI never touched your calendar directly. It asked the server, the server did the work, and you approved the change. That is the whole trick.

Several servers can work together. The official MCP guide walks through a trip to Barcelona: a travel server searches flights and hotels, a weather server checks the forecast, and a calendar and email server blocks the dates and tells your team. One request, three servers.

Local or remote: two kinds of server

The difference is just where the server program runs. Both work the same way once they’re added.

New in 2026, for the curious: the MCP version dated July 28, 2026 made servers stateless. Each message now carries everything the server needs, so the server doesn’t have to remember the conversation between messages. For you, nothing changes. For the people who build servers, they’re now easier and cheaper to run online, which means more remote servers you can simply paste in.

Add your first MCP server

Pick the path that matches your app. Each one takes about five minutes. The menu names below are from the official guides as of October 2026; apps change their menus now and then, so look for “Connectors” or “MCP” if a name has moved.

Path A: a remote server in Claude (easiest, no install)

1. Find the server’s address. The service lists it in its own help pages; it looks like https://…/mcp. To practice, the MCP project offers a test one: https://example-server.modelcontextprotocol.io/mcp.
2. Open Settings. In Claude in the browser, click your profile icon, then Settings. In Claude Desktop, press Ctrl+, (Windows).
3. Click Connectors, then Add, then Add custom connector.
4. Paste the address and click Add.
5. Sign in when asked. This is the service checking it’s really you.
6. Choose what it may do. Back in Connectors, click the new connector to switch individual tools on or off.

Path B: a local server in Claude Desktop (your own files)

This gives Claude access to folders you choose, using the official Filesystem server.

1. Install Node.js (the LTS version) from nodejs.org. To check, open a terminal and type node --version.
2. In Claude Desktop, open the Claude menu, then Settings, then Developer, then Edit Config. This opens a file called claude_desktop_config.json.
3. Replace its contents with this, putting your own Windows user name where it says username. Only list folders you’re happy for Claude to read and change:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "C:\\Users\\username\\Desktop",
        "C:\\Users\\username\\Downloads"
      ]
    }
  }
}

4. Save, then quit Claude Desktop completely and start it again.
5. Check it’s there: click the Add files, connectors, and more button at the bottom left of the message box, then Connectors, then Manage connectors. You should see “filesystem”.
6. Try it: “What files are in my Downloads folder?” Claude will ask your permission before each action.

On a Mac the file is at ~/Library/Application Support/Claude/claude_desktop_config.json, and folder paths look like /Users/username/Desktop.

Path C: Claude Code (for coding projects)

One command in the terminal adds a server. A remote one:

claude mcp add --transport http notion https://mcp.notion.com/mcp

Then type /mcp inside Claude Code to see if it’s connected and to sign in. claude mcp list shows everything you’ve added, and claude mcp remove notion takes it away again.

Using another app? ChatGPT, Visual Studio Code and Cursor support MCP too. Look for “MCP” or “Connectors” in their settings; the server address or settings line is the same idea.

Stay safe: a five-point checklist

An MCP server gets real access to real things, so treat adding one like giving someone a key.

1. Only add servers from sources you trust: the company behind the tool, or the official MCP servers list. A random server from a forum is a stranger with a key.
2. Give the smallest access that works. One folder, not your whole drive. Read-only, if the server offers it.
3. Read every approval request. “Allow delete_file?” deserves a second look. Don’t switch on “always allow” for tools that change or delete things.
4. Be careful with servers that read the web or your email. That content can contain hidden instructions aimed at the AI. This is called prompt injection, and Claude Code’s own docs warn about it.
5. Remove servers you no longer use. Fewer keys, fewer worries.

Never paste passwords or secret keys into the chat. When a server needs a key, the setup puts it in the settings file or the sign-in screen, not in a message to the AI.

Common mix-ups, cleared up

Start with one. Add a single server for the tool you use most, ask it three small questions, and see how it feels. If you also work with AI agents on projects, the guide to giving AI agents context is a good next read.