universal-mcp-fs

A secure, local MCP server that gives AI assistants controlled access to your filesystem.

🛠️ Built With

Overview

universal-mcp-fs gives AI assistants — Claude Desktop today, and any other stdio-compatible MCP client as they add support — filesystem access and command execution on your machine, gated behind an interactive approval system.

• stdio only. No HTTP server, no open ports, no internet exposure.

• Elicitation-based approval. Dangerous actions (delete, run commands, move files) pause and ask the connected client to show a native approval popup — the same kind of Allow/Deny prompt you already see for other tool calls in Claude Desktop.

• Sensitive paths are blocked outright (.ssh, .aws, .gnupg, browser credential stores, /etc/shadow, etc.) — before any approval prompt is even offered.

• 17 tools covering file read/write/move/copy/delete, directory listing, filename and content search, and shell command execution (foreground and background).

Related MCP server: Remote Server MCP

Install

Option 1 — npm (recommended, no build step)

npm install -g universal-mcp-fs

Option 2 — from source (GitHub)

git clone https://github.com/Prabhas125/universal-mcp-fs.git
cd universal-mcp-fs
npm install
npm run build

Setup: Claude Desktop

Edit your Claude Desktop config file:

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• Linux: ~/.config/Claude/claude_desktop_config.json

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

If you installed via npm (Option 1):

{
  "mcpServers": {
    "filesystem": {
      "command": "universal-mcp-fs",
      "args": ["--allowed-dirs", "/home/user;/home/user/projects"]
    }
  }
}

If you built from source (Option 2):

{
  "mcpServers": {
    "filesystem": {
      "command": "node",
      "args": [
        "/full/path/to/universal-mcp-fs/dist/index.js",
        "--allowed-dirs",
        "/home/user;/home/user/projects"
      ]
    }
  }
}

On Windows, use double-escaped backslashes in the path, e.g. "C:\\Users\\yourname\\projects\\universal-mcp-fs\\dist\\index.js".

Then restart Claude Desktop. It will spawn the server automatically and the 17 tools will be available in chat. No further setup, no login, no token.

Verify it's connected

Ask Claude something like "list the files in [one of your allowed directories]". If it responds with a directory listing, the server is connected. If Claude says it has no matching tool, double check the config file path and JSON syntax, then fully quit and reopen Claude Desktop (a reload isn't enough — it only reads this file on startup).

Configuration options

All options can be passed as CLI args in the config's args array, or as environment variables in an env block.

Tool reference

Filesystem — read_file, read_file_lines, write_file, create_directory, list_directory, move_file, copy_file, delete_file, delete_directory

Search — search_files (by filename/glob), search_content (grep-like, with context lines)

Commands — run_command, run_command_background, list_processes, kill_process

Info — file_info, system_info

Full parameter docs are visible to the AI client automatically (and to you, via npx @modelcontextprotocol/inspector).

Permission system

Some tools always require your approval before running: delete_file, delete_directory, run_command, run_command_background, move_file, kill_process, and write_file when overwriting an existing file.

When one of these is called, the server sends an elicitation request to Claude Desktop, which renders it as a native approval dialog — the same UI you already see for regular tool-call confirmations. You can:

• Approve — the action runs once

• Approve + "always allow" — this exact action (same tool, same file/command) is silently approved from then on, persisted to ~/.universal-mcp-fs/always-allow.json

• Decline/cancel — the action is aborted, nothing happens

If you don't respond within 2 minutes, the request automatically resolves to denied — it will not hang the connection or leave Claude waiting indefinitely.

Sensitive paths (.ssh, .aws, .gnupg, browser credential files, /etc/shadow) are rejected before any approval prompt is offered — no amount of clicking "allow" gets past this list. You can extend it via config/default.json if you build from source.

Every decision (approved, declined, always-allow, timed out, blocked) is logged to ~/.universal-mcp-fs/audit.log and to stderr, visible in Claude Desktop's MCP server logs.

Security model

• No network transport — this server never opens a port or listens for external connections.

• All file operations are validated against --allowed-dirs; anything outside is rejected regardless of approval.

• Symlinks are resolved before validation, so a symlink inside an allowed directory can't be used to escape it.

• Path traversal (../..) is normalized away before any check runs.

• If the connected client doesn't support elicitation, dangerous tools fail closed (denied), never fail open.

Publishing / contributing

npm run build
npm version patch     # bumps package.json, creates a git tag
git push && git push --tags
npm publish

Issues and PRs welcome.

License

MIT