Home / Directory / MCP reference servers
Official MCP reference servers: examples, not production connectors
This used to be mistaken for "the" MCP app store. The README now says the opposite: a handful of reference servers, plus a pointer to the MCP Registry for published ones.
What it does
Active reference servers include Everything (a test kitchen), Fetch, Filesystem, Git, Memory, Sequential Thinking and Time. They exist to show SDK usage and protocol features. TypeScript servers can be run with npx.
Older popular servers (GitHub, Slack, Postgres, Puppeteer and others) were archived to modelcontextprotocol/servers-archived. Some have official replacements elsewhere. If a 2025 blog post tells you to npx the old GitHub server, check this README before you do.
The maintainers warn that these examples are not production-ready and that you must apply your own threat model. Filesystem access in particular is a sharp edge.
Who it is for
People learning to build or debug MCP. Also useful when you need a tiny Fetch or Time server in a test harness.
Not the place to get a supported GitHub or Slack integration. Those belong to the vendors now.
Our take
The editorial cleanup is the useful part. The protocol org is no longer pretending to maintain every connector. Reference servers plus a registry is the healthier split.
Limits: GitHub showed no clear SPDX license when we checked. Some examples are still copied into production because they are easy. Filesystem and Git servers can mutate a machine. Sequential Thinking is a prompting toy, not a database.
Read the warning box in the README. It is there because people skipped it.
Compared with AgentKit Drawer
No AgentKit pack is an MCP server. Our Agent Builder pack is system prompts and evals for agents you host, which is a different layer than giving Claude a Fetch tool.
No pack CTA. If you are here to learn MCP, stay in this repo and the SDKs. If you are here to stop an agent breaking your app, go to a stack pack or a free guide.