English · Русский (RU)

🧠 pentool-mcp-server

MCP server for Pentool — the local AI assistant for web pentesting

A self-contained stdio JSON-RPC 2.0 server that brings AI capabilities to any MCP client — first of all to Pentool (BYO-LLM: picking checks, bypassing WAF, finding non-obvious endpoints). Installed with uv in one command.

Why a separate package

MCP is the open protocol for connecting LLM models to tools. pentool-mcp-server encapsulates Pentool's MCP layer as a reusable PyPI package: attached locally (stdio), no network, no ports, installs in seconds, no heavy dependencies.

🔗 Built for Pentool

This server is the MCP component of Pentool, a professional web-security testing terminal (Burp-compatible proxy, scanner, spider, intruder — all in a TUI). It powers Pentool's AI:

• 🎯 picking the relevant scan checks for a concrete target;

• 🛡 WAF bypass / payload suggestions;

• 🕷 finding non-obvious endpoints during spider crawling.

Try the full stack: uv tool install "pentool[ai]" → pentool ai setup → pentool.

⚡ Quick start (uv)

# Install uv (if not present): https://docs.astral.sh/uv/
curl -LsSf https://astral.sh/uv/install.sh | sh

# Standalone — just the MCP server
uv tool install pentool-mcp-server
pentool-mcp-server --version

# Or together with Pentool and its AI extras
uv tool install "pentool[ai]"

Health check (stdio):

echo '{"jsonrpc":"2.0","id":1,"method":"ping"}' | pentool-mcp-server
# → {"jsonrpc": "2.0", "id": 1, "result": {"status": "ok"}}

uv is the standard install path — isolated environment, exactly how Pentool itself is installed. It keeps the server out of your system Python.

🚀 Deploy on a VPS / in a container

The server opens no ports and runs as a local subprocess — safe on any host. With uv it stays out of the system Python:

uv tool install pentool-mcp-server

Full user guide — docs/GUIDE.md. Russian — README.ru.

🧩 MCP tools

🔒 Security

• No network by default. pentool-mcp-server listens only on stdio (stdin → stdout) inside the local process. External access is impossible — neither from other processes nor from another host.

• Do not expose it over TCP/0.0.0.0 without auth. If run on a network port outwardly, anyone able to write to stdin/port gets tools/call generate (resource usage + sending target data to the LLM). MCP has no built-in authentication — bind to 127.0.0.1 and gate via a firewall.

• Target privacy. Data (URL, payload) leaves the host only if an external LLM provider is connected; with a local GGUF model the traffic stays on the machine.

More — docs/GUIDE.md.

🗺 Roadmap

• stdio JSON-RPC 2.0 server (initialize / tools/list / tools/call / ping)

• generate, health, configure tools

• real LLM runner (llama-cpp-python) in model.py

• optional TCP mode with HMAC auth

📄 License

AGPL-3.0 — the same license as Pentool.

Support / bugs: issues · Designed for Pentool · User Guide