Claude Code Desktop and MCP for Beginners

Claude Code can read your files and run commands. MCP lets it also reach your browser, your docs, your issue tracker and your database. That is powerful, and a little scary. This guide shows you how to start in the desktop app, how to add MCP servers step by step, which ones are worth trying for which job, and how not to hand over the keys to your house.

Start Claude Code in the desktop app

Claude Code runs in several places: the terminal, VS Code, the web and the desktop app. The desktop app is the friendliest start, because you get buttons, a diff viewer and a built-in browser pane. The steps below follow the official documentation as of October 2026.

1. Install and sign in. Download the Claude desktop app for macOS, Windows (x64 or ARM64) or Linux (beta). Launch it and sign in with your Claude account.
2. Open the Code tab. The app has three tabs: Chat, Cowork and Code. Click Code.
3. Set up a session. Each conversation is a session. Before you send a message, choose the environment (Local, Cloud, SSH, or WSL on Windows), the project folder, the model and the permission mode.
4. Type a task and press Enter. Start with something safe, like "explain what this project does".
5. Run sessions in parallel. Use + New session (or Ctrl+N / Cmd+N). Turn on the worktree option to give a session an isolated copy of your project (this needs Git).

Permission modes: how much freedom Claude gets

The docs recommend starting complex tasks in Plan mode and switching to Accept edits or Manual after you approve the plan. Other handy parts of the app: the diff view (click the +/- stats to review each changed file and leave comments), the Terminal pane (Ctrl+`), the Browser pane (Ctrl+Shift+B) and a side chat (Ctrl+;) for quick questions that do not clutter the main conversation. The desktop app reads the same settings files as the command-line version.

Do you need the command line too? You will for the claude mcp commands in this post, unless you edit config files by hand. The official quickstart installs it with irm https://claude.ai/install.ps1 | iex in Windows PowerShell, with curl -fsSL https://claude.ai/install.sh | bash on macOS, Linux and WSL, or with winget install Anthropic.ClaudeCode. Then check it with claude --version.

What MCP is, in plain English

MCP (Model Context Protocol) is an open standard that lets an AI app talk to outside tools and data in a common way. Think of it as a universal plug. Claude Code is the client. Each tool you connect, like a browser controller or a documentation service, is an MCP server that offers the AI a list of things it can do (tools), things it can read (resources) and ready-made prompts.

Without MCP, Claude can only work with your files and shell. With MCP, you can say "check the failing test in the browser", "look up the current docs for this library" or "summarize the open issues", and Claude can reach the right tool. In short: you ask, Claude Code sends a request to an MCP server, the server talks to the real tool (a browser, a database, a docs site), and the result comes back to Claude.

Add your first MCP server

There are three kinds of servers, and the transport decides the command:

• Remote HTTP (recommended): a server on the internet. Command: claude mcp add --transport http <name> <url>
• Local stdio: a program that runs on your computer. Command: claude mcp add <name> -- <command> [args] (the -- separates Claude's options from the server's command)
• Remote SSE: an older style. The docs mark it as deprecated, so use HTTP where you can.

Examples from the official documentation:

Scopes: who can see the server

Add --scope project (or -s project) to choose where a server is saved. If the same server is defined in several places, local wins over project, and project wins over user.

Manage your servers

Inside a Claude Code session, type /mcp to check server status and sign in to servers that need OAuth (a browser login). claude mcp list shows a health status next to each server, such as connected, needs authentication or failed to connect.

Writing the command: a cheat sheet

There are only a few shapes. Replace the capitalized parts with your own values, and use placeholders instead of real secrets.

Rules that save you from errors:

• NAME should use only letters, numbers, - and _.
• For a remote server, the URL should start with https://.
• For a local server, the -- separates Claude's options from the server's own command. Put the server command after it.
• When you use --env, put at least one other option (such as --transport) between it and the name. The docs warn that otherwise the name can be read as another KEY=value pair.
• Never type a real token into a command that lives in a shared note or chat. Use placeholders, and set the real value in your own environment.

With project scope, the server is saved in a .mcp.json file in your project root, which your team can share. This is what the first row above looks like there (a URL entry needs "type": "http", or it is treated as a configuration error):

And a local server that needs a key. The ${...} form reads the value from your own environment, so the secret never sits in the file:

Useful MCP servers by area

There are thousands of MCP servers, and many are community projects. Instead of a giant list, here is a short, practical map by job. Every command below comes from the official Claude Code documentation or the server's own project page. Always check the project's page for the current command before you run it.

What each kind of MCP server is for.

Lower riskMediumBe carefulTap an area to filter the table below.

The servers, in one table

Pick an area to filter the list. Risk is my own rough judgment, not an official score.

Commands come from the official Claude Code documentation or the project's own page. Always check the page for the current command before you run it.

Is this server safe to install?

An MCP server is code that runs with your permissions or talks to your accounts. The official docs give a plain warning: verify you trust each server before connecting it, because servers that fetch outside content can expose you to prompt injection (hidden instructions in web pages or documents that try to steer the AI). Claude Code also asks for your approval before using project-scoped servers from a .mcp.json file, and a cloned repository cannot approve its own servers.

Use this checklist before you install a server you do not know. The more answers you have in the right-hand column, the more careful you should be.

What to do about each risky answer:

• Unknown maker: find out who made it, and prefer the official server of the service you already use.
• Closed source: you cannot read the code, so rely on the vendor and keep permissions small.
• It can change data: start with read-only access if the server offers it.
• It can delete data or move money: use a test account, require approval for every action, and never use it on production first.
• It reads outside content: that content can carry hidden instructions (prompt injection). Do not give the same server strong write access at the same time.
• Full-access key: use a limited token or a test key instead.
• Team scope: project scope shares the server through .mcp.json. Review it, keep secrets in variables, and expect each person to approve it.

Troubleshooting and tips

• Server will not connect. Run claude mcp list. The status line shows connected, needs authentication or failed to connect, and claude mcp get <name> shows the failure detail. In a session, /mcp shows status and lets you sign in.
• Local servers need their runtime. Many stdio servers start with npx (needs Node.js) or uvx (needs uv). Check that the command works in your own terminal first.
• Slow start. The startup wait can be changed with the MCP_TIMEOUT environment variable (in milliseconds).
• Big tool outputs. There is a default limit on how much a tool result can contain. The docs describe the MAX_MCP_OUTPUT_TOKENS variable for raising it.
• Too many servers? You do not need to worry much: Claude Code uses tool search, which only loads tool names at the start and fetches full definitions when needed.
• In the desktop app, Connectors are MCP servers with a point-and-click setup (use the + next to the prompt box, or Settings, then Connectors). For anything not listed there, add the server in your settings files, as shown above.
• Start small. Add one server, try a real task, then decide if you need another.

FAQ

Disclaimer: commands and UI names follow the official Claude Code documentation as checked in October 2026 and can change, so please check the current docs. Third-party MCP servers are not reviewed or endorsed by this blog, and the risk ratings are my own rough judgment. Never paste real passwords or keys into shared files or chats. Use test data and read-only access when you try something new.