If you're running the official MCP filesystem reference server (modelcontextprotocol/servers) on Windows and something's failing in a way that feels like it should just work, there's a real chance it isn't your config. As of November 2026 there are four separate open issues against that server that only reproduce on Windows β and none of them has a merged fix.
This is the kind of thing worth checking before you spend an hour reinstalling, re-pathing, or re-reading your client config. Five minutes matching your exact symptom against these four issues will tell you whether you're debugging your own setup or waiting on an upstream fix.
#5180 β backslash home-path expansion fails. A home-directory path written with native Windows backslashes doesn't expand correctly. If your config uses ~\Documents\... instead of a forward-slash or already-expanded path, this is likely you.
#5185 β realpath throws ENOENT on SUBST or mapped drives. If you're pointing the server at a SUBST-mapped or network-mapped drive letter, realpath resolution can throw ENOENT even though the path is valid and reachable from every other tool on the machine.
#5260 β parent-directory-swap security gap. An earlier attempt to close a directory-escape hole left it only partially closed. A crafted parent-directory path can still, under specific conditions, escape the server's intended root boundary on Windows. Treat this as a reason to scope the server's allowed directories conservatively, not just a bug to wait out.
#5375 β NTFS alternate-data-stream bypass (newest). A path like file.txt:stream β NTFS's alternate-data-stream syntax β slips past path validation that was written assuming a plain filename. Like #5260, this is a validation bypass with real security implications, not just a broken feature.
Match the symptom, not the error message, since the error text is often generic:
ENOENT on a drive letter that works everywhere else β #5185:streamname it shouldn't β #5375
If none of those match, it's more likely your own config β worth ruling the four out first precisely because "upstream bug, no local fix" and "my config is wrong" require completely different next steps. The practical mitigations right now: use forward-slash or already-expanded paths, avoid SUBST or mapped drives as the server's working directory, and don't treat this server's path validation as a hard security boundary on Windows until #5260 and #5375 are fully resolved. None of that patches the server β only an upstream merge does, and none has landed as of this writing. Check the issue numbers directly rather than trusting any cached summary, this one included, since the moment one merges this is stale.
If your actual problem is "the tools don't show up at all" rather than one of these four specific Windows symptoms, that's a different, more common failure β usually process, transport, handshake, or client-registry related, not server-side. We maintain a free reference for that broader diagnosis order (the MCP Debug Kit), but it's worth saying plainly: it doesn't and can't patch any of the four bugs above. Nothing client-side can patch an upstream server. What it's for is confirming which kind of problem you actually have before you spend more time on it β which, for these four, means stopping local debugging and going to track the GitHub issue instead.