Everyone who has watched an agent read a lockfile knows the problem this addresses: the reference server returns whatever it finds, and a single call can consume the entire context. Capping reads, listings and traversal depth is unglamorous and immediately useful. The directory allowlist remains the important setting — pass the narrowest set of paths that works.
A fork of the reference filesystem MCP server, changed in one direction: making responses small enough for a model to use. It reads and writes files, creates and navigates directories, searches, and edits existing files, with every operation restricted to the directories you name on the command line.
- File and directory operations: read and write files, create and navigate directories, search for files, and make precise edits to existing ones
- Limits on file reading, so an enormous file is not returned in full
- Caps on directory listings, so a folder with thousands of entries still returns something usable
- Depth-limited tree traversal, so a deep folder structure does not consume the whole response
- Search results trimmed to the relevant matches instead of every hit
- Automatic skipping of large dependency folders such as `node_modules` and `.git`
- A hard boundary: only the directories passed as arguments are reachable, attempts outside them are rejected, and all file operations are logged
Nothing — no account, no key. The documented client entry runs `npx github:codemaestroai/filesystem-mcp` with one or more absolute project paths as arguments, over stdio. The package is `codemaestroai-filesystem-mcp` (0.6.2 in package.json). The publisher notes this is a temporary measure while native file operations are built into their desktop app.
One command — npx github:codemaestroai/filesystem-mcp <YOUR_PROJECT_PATH>