Connect Claude Code to Your Tools with MCP: The Tutorial

You've installed Claude Code, it reads your project, it writes code. But when you ask about the status of your last deployment or the contents of a ticket, it can't answer: all it sees are the files in your folder. MCP (Model Context Protocol) solves exactly that problem. It's an open standard published by Anthropic in late 2024 — a sort of universal plug that lets an AI assistant talk to your tools: GitHub, Sentry, Notion, a database, a browser, your file system. You hook up an MCP server once, and Claude Code knows how to use it in every session that follows.

This tutorial gives you three paths to get there: a guided track if you're truly starting from scratch, Claude Code's native command (two minutes, built right into the tool), and the shared config file for when you work on a team. No heavy technical prerequisites, and every term gets explained along the way.

What exactly is MCP?

MCP is an open standard that lets Claude Code use external tools (APIs, databases, applications) through small programs called MCP servers.

The simplest analogy: before USB, every device had its own cable. With MCP, every tool exposes its capabilities in a single format, and any compatible assistant can plug into it. Anthropic released the protocol as open source in November 2024; it has since been adopted by other vendors, which means an MCP server written today works beyond Claude Code.

Four pieces of vocabulary are enough to follow the rest:

• MCP server: the small program that exposes a tool. Example: the Sentry server knows how to read your production errors.
• MCP client: the software that consumes those servers. Here, that's Claude Code.
• Tools: the actions the server permits (create a ticket, run a SQL query, open a web page).
• Resources: the data the server makes available for reading, which you can call in a message using the at-sign syntax, for example @sentry:issue.

On top of that comes the idea of transport, meaning how Claude Code talks to the server. There are three options: stdio (the server runs on your machine), SSE and HTTP (the server is hosted remotely by the tool's vendor). You don't have to guess: each service states in its documentation which transport and which address to use.

In practice, once a server is connected, you can write to Claude Code: "look at the last three Sentry errors on the app, find the cause in the code, and propose a fix." It goes and gets the information itself instead of waiting on you to copy and paste.

Method 1: Learn MCP in a ready-made environment with Skilzy

If you've never opened a terminal, the fastest route is to practice in a guided setting, with the tools already installed, rather than configuring your machine blind.

That's the logic behind Skilzy, a French e-learning platform covering more than 15 programs for learning and using AI: creating images, videos (UGC, ads, faceless channels, avatars, voiceovers), music, automating with n8n, writing good prompts, coding with AI. The Claude Code and the terminal program is the one you want here: it starts with installation, moves through the basic commands, the project instructions file, hooks, and then connecting MCP servers.

The main draw for a beginner is the built-in AI Lab: you use the real tools directly inside the platform, credits included, without stacking up five subscriptions at 20 euros a month just to experiment. You learn the logic of the protocol and reproduce the steps while someone explains what each line does.

Worth knowing as well:

• Two state-recognized, fundable certifications are available: RS7439 (content marketing with AI) and RS6792 (AI and sales). You can verify they exist on the public France Compétences registry.
• There's a B2B offer for training organizations that want to integrate these tracks.

Be clear-eyed about one thing: a course won't configure your machine for you. At some point, you'll be typing the commands yourself. The next two methods spell them out in full, and they're free.

Method 2: The claude mcp add command (Claude Code's native feature)

This is the official and fastest route: a single command in your terminal is enough to hook up a server, with no extra software.

Two prerequisites. First, Claude Code has to be installed on your machine; if that's not done yet, follow the guide to install Claude Code on macOS or its Windows equivalent. Second, many local servers ship as Node.js packages, so Node.js version 18 or higher needs to be present (type node --version to check).

Hook up a local server (stdio)

The server runs on your machine. The double dash separates Claude Code's options from the command to run:

claude mcp add files -- npx -y @modelcontextprotocol/server-filesystem ~/projects

Hook up a remote server (HTTP or SSE)

Here you don't supply a command but an address, provided by the service's vendor:

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

claude mcp add --transport sse linear https://mcp.linear.app/sse

Most of these remote servers require authentication. Launch Claude Code, type the /mcp command in the session, pick the server, and follow the sign-in flow in your browser. The token is stored by Claude Code — you don't copy it anywhere.

Choosing the scope

This is the point beginners miss most often. A server isn't automatically available everywhere.

Scope	Command	Available where	What for
local (default)	claude mcp add ...	Only in the current project, for you	Testing, experimental servers
project	claude mcp add --scope project ...	For anyone who clones the repo	Tools shared across the team
user	claude mcp add --scope user ...	In all your projects, for you alone	Your personal day-to-day tools

Example of a server you want everywhere:

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

Managing what you've connected

claude mcp list
claude mcp get sentry
claude mcp remove sentry

The first command shows the connection status of each server, the second gives the details of one server, the third deletes it. These three naturally join the Claude Code commands you absolutely need to know in your daily workflow.

Method 3: The .mcp.json file for the whole team

When several people work on the same repo, you declare the servers in a .mcp.json file versioned alongside the code, and everyone picks them up automatically.

This file goes at the root of the project. It's exactly what the command with the project scope generates, but writing it by hand gives you more control:

{
  "mcpServers": {
    "postgres": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres", "postgresql://localhost/my_database"],
      "env": {}
    },
    "sentry": {
      "type": "http",
      "url": "https://mcp.sentry.dev/mcp"
    }
  }
}

Non-negotiable security rule: no API keys, no passwords in plain text in this file, since it goes into Git. Claude Code can read your machine's environment variables using the curly-brace syntax:

"env": { "API_KEY": "${MY_API_KEY}" }

Each developer then sets MY_API_KEY on their own side, and the repo stays clean. Also make sure your .gitignore properly covers your environment files.

When someone opens the project for the first time, Claude Code asks for confirmation before enabling the declared servers: that's a safeguard against a malicious repo. If you answered wrong, the claude mcp reset-project-choices command resets those approvals.

This file pairs well with your project instructions file and your automations: if you want to go further, the tutorial on Claude Code hooks shows how to trigger automatic actions around these tools.

Checking that it works, debugging, and staying safe

Type /mcp in a Claude Code session: you immediately see which servers are connected, which are failing, and which are waiting on authentication.

The most common failures and their real causes:

• The server stays in "failed": the launch command can't be found. Check that Node.js is installed and that the package name is exact. Restart Claude Code with the --debug option to read the full error message.
• The server takes too long to start: increase the startup timeout, for example with MCP_TIMEOUT=15000 in front of the claude command (the value is in milliseconds).
• The server is connected but Claude isn't using it: explicitly ask for the tool in your message ("use the sentry server to..."). You can also spell it out in your CLAUDE.md file so it becomes automatic.
• On Windows, an npx command fails: sometimes you need to wrap it in cmd /c. The official MCP documentation for Claude Code covers this case.

On the security side, three habits to adopt from your very first server. One: an MCP server is third-party code running on your machine with your permissions, so only install servers whose publisher you know. Two: apply the principle of least privilege — for example, connect your database with a read-only user rather than the admin account. Three: watch out for prompt injection, meaning content fetched by a server (a ticket, a web page, a comment) that contains hidden instructions aimed at the AI. Always re-read the sensitive actions Claude Code proposes before approving them, especially when they involve external data.

What you can do in the next hour

Start small: connect a single server, ideally for the tool you open ten times a day. Check with /mcp that it responds, ask Claude Code two or three questions to see what it can pull out, then add the next one. Once three servers are running, switch to the project scope and share the configuration with your team. If terminal commands are still holding you back, a guided track will save you weeks of trial and error; otherwise, the official documentation and this tutorial are more than enough to get going.