How to Set Up MCP Servers for Claude, Step by Step

The short answer (July 2026): MCP servers connect Claude to your real tools: files, GitHub, Notion, databases, error trackers. In Claude Code, one command does it: claude mcp add --transport http notion https://mcp.notion.com/mcp. In Claude Desktop, you edit one JSON config file and restart the app. Both methods take under five minutes, and this guide gives you the exact commands, the three servers worth installing first, and the security rules the official docs insist on.

Last verified: July 10, 2026, against modelcontextprotocol.io and Anthropic's Claude Code docs.

What MCP is, in one minute

The Model Context Protocol is an open standard for connecting AI applications to external systems. The protocol's own analogy is the best one: "Think of MCP like a USB-C port for AI applications." One plug shape, any device. A server you set up once works across Claude, and the standard is bigger than one vendor: per the protocol's site, ChatGPT, VS Code, and Cursor support it too, and the project is governed under the Linux Foundation's LF Projects umbrella with maintainers holding seats as individuals, not companies.

The practical payoff: instead of pasting context into a chat, Claude reads your actual Notion pages, queries your actual database, and files your actual GitHub issues, with your approval gating each action.

If "agent" is still an abstract word for you, start with what is an AI agent; MCP is how agents reach the real world.

Setup path 1: Claude Code (the one-liner)

Claude Code speaks three transports. You will mostly use the first.

Remote HTTP servers (the modern default, used by Notion, GitHub, Sentry, Stripe and most SaaS tools):

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

Servers needing a token take a header: --header "Authorization: Bearer your-token". Servers using OAuth, like Sentry, authenticate in-session: run /mcp and follow the sign-in.

Local stdio servers (run on your machine, for local resources):

claude mcp add --transport stdio airtable --env AIRTABLE_API_KEY=YOUR_KEY -- npx -y airtable-mcp-server

The -- matters: everything after it is the server's own command, everything before it is Claude's flags.

Scopes decide who gets the server. Add -s user for all your projects, -s project to share it with your team via a .mcp.json checked into the repo (Claude Code asks each teammate for approval before using it), or leave the default local scope for just this project.

Housekeeping: claude mcp list shows what's installed, claude mcp remove <name> uninstalls, and /mcp inside a session shows live status. If you already configured servers in Claude Desktop, claude mcp add-from-claude-desktop imports them on macOS and WSL.

Setup path 2: Claude Desktop (the config file)

For the desktop app, per the official quickstart:

1. Open the Claude menu in your system menu bar, then Settings, then the Developer tab, then Edit Config.
2. That opens claude_desktop_config.json (on macOS under ~/Library/Application Support/Claude/, on Windows under %APPDATA%\Claude\).
3. Add a server. The Filesystem server, which lets Claude read and write chosen folders, looks like this:

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

4. Fully quit and restart Claude Desktop. Verify under the "Add files, connectors and more" control in the chat input.

You need Node.js installed for npx-based servers. Every file action still requires your explicit approval per operation.

Which servers to install first

Beyond these, two directories are worth bookmarking: the MCP Registry ecosystem listing, and Anthropic's reviewed directory at claude.ai, which Claude Code can install from directly. One landscape note: several early first-party servers (Slack, PostgreSQL, Brave Search) are now archived or maintained by third parties, so check a server's maintenance status before depending on it.

The security rules (not optional)

MCP servers run with real access to your stuff, and the official guidance is blunt. Three rules cover most of the risk:

• Trust before you connect. Claude Code's docs: "Verify you trust each server before connecting it. Servers that fetch external content can expose you to prompt injection risk." Anthropic reviews its directory listings but "does not security-audit or manage any MCP server."
• Scope access down. The Desktop quickstart's warning applies everywhere: the server "runs with your user account permissions, so it can perform any file operations you can perform manually." Grant the one folder, not the home directory; the read-only token, not the admin one.
• Keep approval prompts on. Per-action approval is the safety net that catches a compromised or misbehaving server. Auto-approving everything to save clicks removes exactly the check that matters.

Troubleshooting

Server not showing up in Claude Code: run /mcp to see status and errors, and claude mcp get <name> to inspect its config. In Claude Desktop, the classic fix is real: fully quit the app (not just close the window) and relaunch; config only loads at startup. Auth failures on remote servers are usually an expired token in the --header flag or an OAuth session that needs /mcp sign-in again.

FAQ

Do I need to know how to code to use MCP?
No for remote servers in Claude Desktop and Claude Code: it's one command or one config block, copy-pasted. Building your own server is a coding task, and Anthropic Academy has a free Introduction to Model Context Protocol course for that.

MCP servers vs Claude connectors: what's the difference?
Same protocol, different packaging. Connectors are Anthropic's user-friendly, reviewed wrapper for claude.ai and the apps; claude mcp add is the direct route with full control over transports and scopes. Power users end up using both.

Does MCP only work with Claude?
No. It's an open standard under the Linux Foundation, and ChatGPT, VS Code, and Cursor support it, per modelcontextprotocol.io. A server you adopt today isn't vendor lock-in.

What should my first MCP project be?
Connect Filesystem to one working folder and GitHub to one repo, then give Claude a task that needs both, like "read the TODOs in this project and file them as GitHub issues." Ten minutes, and you'll understand viscerally what agents change. Our Claude Code guide covers the tool this all plugs into.

The MCP ecosystem moves monthly, servers get archived, transports get deprecated; this page is re-verified as it shifts. Track the ecosystem as news with AI Weekly, free, 3 issues a week, read by 40,000+ practitioners.