MCP

The Best MCP Servers to Start With

4 min read

Start with the reference servers that are still maintained, not the ones every old blog post lists. The modelcontextprotocol/servers repo now keeps seven: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time.

1. The maintained shortlist

2. Check the archive before you install

This is the part that trips people up. GitHub, PostgreSQL, Puppeteer, Slack, Google Drive, Google Maps, GitLab, Redis, Sentry, SQLite and Brave Search have all moved to the servers-archived repository. They still install, but nobody is maintaining them. Where a first-party replacement exists, use it: GitHub is now a hosted server at https://api.githubcopilot.com/mcp/, and Brave ships @brave/brave-search-mcp-server.

3. Add one

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

What it does: registers a stdio server; everything after -- is the command Claude Code runs to launch it, and the trailing path is the only directory it may touch.

4. Confirm it connected

/mcp

What it does: lists every configured server with its health — ✔ Connected, ! Needs authentication, or ✘ Failed to connect — plus its tool count.

If it fails

✘ Failed to connect on a fresh install is almost always the launch command, not the server — run the part after -- in your own shell and read the error. Servers load at session start, so a change needs a new session.

Give it the least access that works

An MCP server runs with your permissions, and the text its tools return lands in your prompt. Scope a token to one repo, use a read-only database user, allow a single folder. Treat tool output as data: a web page should never tell Claude to run a command.

Full walkthrough: install an MCP server. For GitHub, see connect Claude Code to GitHub.

Source: modelcontextprotocol/servers.

Open the full interactive version (with copy buttons) ↗

← All guides