MCP Server Picker

What an MCP server actually is

MCP is the Model Context Protocol - a small, open standard for letting Claude Code talk to systems outside your filesystem. Without MCP, Claude Code can read your code, run your scripts, and edit your files. That is already a lot, but everything it knows about your project ends at the directory boundary. With MCP, you can give the model controlled access to your database, your GitHub repos, your Notion docs, your project tracker, or a real browser. Each MCP server is a small process that exposes a specific surface area; the model calls into it the same way you would call into a library.

The reason MCPs matter so much is that ninety percent of the friction in real software work is not in the code itself - it is in the things outside the code. The schema you keep forgetting. The ticket you keep tabbing back to. The pull request you keep re-reading. MCPs collapse that context into the same session you are coding in. The first time you ask Claude Code 'how many users signed up yesterday' and it actually queries your database and tells you the answer, you understand why this protocol exists.

Why three and not thirty

There are dozens of MCP servers in the public catalog. Most projects need three. The temptation when you discover MCPs is to install every one that looks interesting, but each MCP that is loaded into a session costs context, costs setup time, and gives the model more surface to wander into. The healthy pattern is: install the two or three MCPs that match your daily work, use them for a week, and only add more when you have a specific job they would do.

• If you build SaaS - the database MCP is almost always the highest-leverage install.
• If you do client work or open source - the GitHub MCP turns repo reading into a one-line ask.
• If you work in fast-moving libraries - the web search MCP keeps the model grounded in current docs instead of stale training.
• If your work touches a browser - Puppeteer / browser MCPs let the model verify what it just shipped.
• If your team lives in Notion / Linear / Slack - those MCPs collapse a tab into a tool call.

How to install (the actual steps)

1. Pick your top one (the calculator above does this for you).
2. Run the install command from the result card. Most are a one-liner: `claude mcp add <name> <command>`.
3. Supply any required environment variables in your shell or in a local .env.
4. Restart your Claude Code session. The new MCP appears in the available tools list.
5. Test it with a small, low-stakes question. If it returns clean output, you are in.

What MCPs are NOT for

MCPs are not a substitute for good prompts. They are not a substitute for a clean CLAUDE.md. They are a substitute for tab-switching - for the moments when you would otherwise stop, open another tool, copy something, paste it back, and continue. If you find yourself installing an MCP because you hope it will fix bad output, the problem is upstream of the MCP. Fix the prompt first, fix the CLAUDE.md second, and only then reach for a new MCP.

Where MCPs fit in the broader workflow

The Claude Code Club curriculum has a full phase on MCPs - from installing your first one to building your own from scratch when no off-the-shelf MCP exists. The picker above is the front door: it gets you running with three MCPs that match your actual work in under ten minutes. The deeper learning is what turns MCPs from 'thing I installed' into 'core part of how I ship.'