MCP Server Picker

The MCP ecosystem grows fast and unevenly. Anthropic maintains a tier of official servers. A growing community publishes their own — some excellent, some abandoned three commits after launch. Users routinely waste hours on the wrong package. This skill's job is to cut that waste with opinionated, tiered recommendations rather than a neutral "here are your options" list.

The first move is the most useful: check whether MCP is needed at all. Claude Code already does git, file reading, file writing, bash, and curl natively. Half the time someone asks for an MCP server, they're describing a capability Claude Code already has. Naming that explicitly saves installs that would have added latency and broken nothing useful.

Recommendations come in tiers. Tier 1 is official or battle-tested — GitHub, Filesystem, Postgres, SQLite, Brave Search, Memory, Slack, Google Drive, Puppeteer, Fetch — recommended with confidence. Tier 2 is community servers that work with caveats, paired with an evaluation rubric (last commit date, open-issue volume, weekly downloads, README quality) so you can judge them yourself. Tier 3 is build-your-own with a pointer to the MCP SDK when nothing fits.

The skill is opinionated about what to avoid. Don't stack ten servers — each is a process. Don't use community servers for sensitive data without auditing the code. Add a human-in-the-loop checkpoint for anything that sends emails, posts to channels, or modifies production data; start read-only and add writes deliberately. Don't pre-install "just in case" — add a server the first time you hit a task Claude can't do without it. The starter sets at the end are role-specific (developer, researcher, everyone else) rather than the maximalist list most guides default to.