📑 Table of Contents

• Overview
• Features
• Demo
• Tech Stack
• Requirements
• Quick Start
• Configuration
• Packaging as a Desktop Extension
• How It Works
• Project Structure
• Documentation
• Security & Privacy
• Contributing
• License
• Author

🔭 Overview

Every time an LLM reads a whole file just to answer a small question, it burns thousands of tokens it didn't need. MCP Filesystem Agent is a Model Context Protocol server that gives Claude 22 purpose-built file operations — preview reads, chunked pagination, AST-based code search, dry-run edits — so it gets exactly the context it needs and nothing more.

The problem: naive "read the whole file" tool-use patterns don't scale past a few hundred lines before they eat your context budget.

The solution: structured, scoped operations — read_file(preview_lines=50) instead of dumping 2,000 lines; search_code_structure() instead of reading five files to find one function; replace_text(dry_run=True) so edits are previewed before they're committed.

Who it's for: developers who want Claude to work directly across a real codebase or document tree — through Claude Desktop, Claude Code, or any MCP-compatible client — without babysitting what gets read.

✨ Features

• 🎯 Scoped by design — access is limited to directories you explicitly allow, passed as CLI arguments exactly like the official filesystem MCP server. Nothing outside that scope is reachable.
• ⚡ Preview & chunked reads — pull the first N lines or a specific byte-range chunk instead of an entire file, so large files don't blow the context window.
• 🔍 Multi-language code intelligence — AST-accurate function/class/import extraction for Python, regex-based for JavaScript, Go, and Rust.
• ✏️ Dry-run edits — preview a find-and-replace before committing it, so you see the diff before anything changes on disk.
• 🧭 Fast search, not full reads — search by filename, extension, or content (plain text or regex) with contextual snippets instead of whole-file dumps.
• 📦 Batch reads with size guards — read multiple related files in one call, capped so a batch can't quietly consume your whole budget.
• 🐳 Production Docker setup — non-root user, resource limits, a real process healthcheck (not a no-op).
• 🧪 Actually tested — path-safety, CLI-arg handling, and content search all have regression tests in tests/.

🖼️ Demo

Once installed, Claude Desktop renders a native config screen generated directly from manifest.json — pick allowed directories, and tools are auto-grouped into "Read-only" vs. "Write/destructive" based on each tool's annotations.

You: What's in my project, and does it have any TODOs left?

Claude uses: get_tree(path=".", max_depth=2)
             search_content(query="TODO")

Result: A directory tree plus every TODO with file and line number —
        without Claude reading a single full file.

🧱 Tech Stack

📋 Requirements

🚀 Quick Start

# 1. Clone
git clone https://github.com/Mdskun/mcp-fs-agent.git
cd mcp-fs-agent

# 2. Install dependencies
pip install -r requirements.txt

# 3. Run — pass the directory Claude should access as an argument
python server3.py /path/to/your/projects

You should see:

============================================================
🚀 MCP FILESYSTEM AGENT v3 (PRODUCTION-READY)
============================================================
📁 BASE DIRECTORIES (1):
   1. /path/to/your/projects
============================================================

Connect it to Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "filesystem-agent": {
      "command": "python",
      "args": ["/path/to/mcp-fs-agent/server3.py", "/path/to/your/projects"]
    }
  }
}

Restart Claude Desktop. Add more allowed directories by listing more paths in args — one server, multiple scoped roots.

Run it in Docker instead

docker-compose up -d
docker-compose logs -f     # watch it start
docker-compose down        # stop it

Prefer a system service?

# /etc/systemd/system/mcp-fs-agent.service
[Unit]
Description=MCP Filesystem Agent
After=network.target

[Service]
Type=simple
User=your-username
Environment="MCP_BASE_DIR=/home/your-username/projects"
ExecStart=/home/your-username/mcp-fs-agent/venv/bin/python /home/your-username/mcp-fs-agent/server3.py
Restart=always

[Install]
WantedBy=multi-user.target

sudo systemctl daemon-reload
sudo systemctl enable --now mcp-fs-agent

nssm install mcp-fs-agent "C:\path\to\venv\Scripts\python.exe" "C:\path\to\server3.py" "C:\Users\YourName\Projects"
nssm start mcp-fs-agent

⚙️ Configuration

Simplest: command-line arguments (recommended)

python server3.py /path/to/projects
python server3.py /path/to/projects /path/to/documents /data/external   # multiple roots

Alternative: environment variables

Priority order: CLI args → MCP_BASE_DIRS → MCP_BASE_DIR → default.

Tunable limits

Edit near the top of server3.py:

MAX_FILE_SIZE_KB   = 2000   # single-file read/write cap (2MB)
MAX_RESULTS        = 50     # max search results returned
DEFAULT_CHUNK_SIZE_KB = 50  # chunk size for read_file_chunked()
TOTAL_BATCH_SIZE_KB = 5000  # cap for batch_read_files()
MAX_LINES_TO_SEARCH = 10000 # cap for search_content()

📦 Packaging as a Desktop Extension

The repo ships manifest.json (MCPB spec 0.3) so it can be packaged as a .mcpb bundle and installed as a Claude Desktop extension — this is what renders the native "Allowed Directories" config screen.

# 1. Bundle the mcp dependency using the *exact* interpreter that will run it
/usr/bin/python3 -m pip install "mcp>=1.9,<2" --target=server/lib

# 2. Install the MCPB CLI
npm install -g @anthropic-ai/mcpb

# 3. Validate, then pack
mcpb validate manifest.json
mcpb pack .

Interpreter mismatch is the #1 failure mode. pydantic_core ships a compiled binary — if you pip install --target=server/lib from a venv or conda environment that isn't the exact Python Claude Desktop launches (usually /usr/bin/python3), you'll get ModuleNotFoundError: No module named 'pydantic_core._pydantic_core' at runtime. Always install with the target interpreter itself (/usr/bin/python3 -m pip install ..., not a bare pip install ...).

Install the resulting .mcpb in Claude Desktop, test it against a disposable directory first, then submit via the Desktop Extension form (a separate path from the Connectors Directory portal, which is for remote HTTPS servers only).

🧠 How It Works

User flow:

1. You point the server at one or more directories (CLI args, env var, or the Desktop Extension's directory picker).
2. Claude calls a tool — say, search_content(query="TODO").
3. Every path is resolved and checked against the allowed directories before any file touches disk.
4. A compact, structured ToolResponse comes back — not a raw file dump.

Internally:

Claude ──▶ FastMCP tool call ──▶ safe_path() validation ──▶ file operation ──▶ ToolResponse
                                       │
                                       └─ segment-aware check via Path.relative_to()
                                          (rejects sibling-directory & ../ traversal)

There's no framework pattern here beyond "one function per tool" — it's a single, flat module by design, prioritizing readability over abstraction for a project this size.

📁 Project Structure

mcp-fs-agent/
├── server3.py              # The entire server — all 22 tools, one file
├── manifest.json           # MCPB packaging manifest (Desktop Extension config)
├── requirements.txt        # Runtime dependency (mcp SDK)
├── .mcpbignore             # Excludes venv/tests/docs from the packed bundle
├── Dockerfile              # Non-root, healthchecked container build
├── docker-compose.yml      # One-command Docker deployment
├── PRIVACY.md              # Data-handling policy (required for submission)
├── SECURITY.md             # Security features, known limitations, disclosure process
├── CHANGELOG.md            # Version history, including real fixes (not just features)
├── CONTRIBUTING.md         # Dev setup, tool template, PR process
└── tests/
    ├── test_safe_path.py       # Path-traversal regression tests
    ├── test_cli_args.py        # CLI-arg config priority tests
    └── test_search_content.py  # Content-search matching regression tests

📚 Documentation

• 📝 CHANGELOG.md — every version, including the real bugs that got fixed (a string-prefix path check, a broken match object in search) — not just a feature list.
• 🔒 SECURITY.md — security model, Docker hardening, and how to report a vulnerability.
• 🔐 PRIVACY.md — what data this touches (nothing leaves your machine).
• 🤝 CONTRIBUTING.md — dev setup, the required tool-annotation pattern, commit conventions.

🔒 Security & Privacy

• Path validation is segment-aware (Path.relative_to()), not a naive string-prefix check — sibling directories that share a name prefix with an allowed folder can't be reached.
• Zero network calls. No telemetry, no analytics, nothing sent anywhere. See PRIVACY.md for the full policy.
• Every tool is annotated (readOnlyHint / destructiveHint) so Claude Desktop can correctly group and gate write/destructive operations from read-only ones.

Full details in SECURITY.md.

🤝 Contributing

Contributions are welcome — especially test coverage for the write/edit tools, which is the biggest known gap right now.

git checkout -b feature/your-feature-name
# make changes, add tests
pytest tests/ -v

See CONTRIBUTING.md for the full dev setup, the required tool-annotation pattern for new tools, and PR conventions.

📄 License

Released under the MIT License — use it, modify it, ship it.

👤 Author

Manthan (@Mdskun)

If this saved you a context window or two, a ⭐ on the repo is the easiest way to say thanks.