Garak MCP Server

Example Use Cases

• 1

  Automate LLM security testing by running predefined Garak probes against specified models and collecting vulnerability findings.

• 2

  Enumerate available model types and installed models (ollama, openai, huggingface, ggml) before launching targeted attacks or diagnostics.

• 3

  Integrate vulnerability scanning into MCP-capable GUIs (e.g., Claude Desktop, Cursor) and retrieve the latest scan report programmatically.

Description

This repository provides an MCP (Model Context Protocol) server to run Garak vulnerability scans and attacks against models (ollama, openai, huggingface, ggml). It exposes tools to list model types and models, enumerate available Garak probes, run attacks, and fetch the last report. It is designed for easy integration with MCP-compatible clients (e.g., Claude Desktop, Cursor) and aims to centralize automated security testing of LLMs. Requires Python 3.11+ and the 'uv' runner; ollama is optional for testing ollama-hosted models.