Reference MCP Servers

The canonical implementations to read before writing your own

The Model Context Protocol organisation maintains a set of reference servers covering filesystem, fetch, git, memory and sequential thinking. They are useful as tools and more useful as the clearest available specification of how a server should behave.

FreeMITTypeScript and PythonSelf hostable

Read these before you build one

Every question about how to shape a tool, name an argument, describe a resource or report an error is answered somewhere in this repository. Reading the filesystem server end to end is a faster education than any amount of protocol documentation.

The filesystem server in particular shows how to constrain a server to specific directories, which is the pattern every server touching a real machine should copy.

As tools in their own right

The fetch server gives an agent web access with a defined boundary. The git server exposes repository history in a structured form. Both are small enough to audit before you trust them, which is not true of most servers you will install.

Note that some servers here are explicitly reference material rather than production software. Read the repository's own guidance on which are archived before depending on one.

Reach for it when

you are writing your own MCP server and want a correct model to copy
you want a small server you can actually audit before trusting it
you need filesystem access scoped to specific directories

Watch out for

some entries are reference material rather than maintained products, so check before depending on one
a filesystem server is only as safe as the directories you scope it to
the fetch server can reach any URL, which is worth thinking about before enabling it

NEXT TO IT ON THE SHELF

Reference MCP Servers

Read these before you build one

As tools in their own right

Reach for it when

Watch out for

Compare with these

Chrome DevTools MCP

Playwright MCP

Context7

New entries, as they land