GuardiaMCP: A Free, Offline, Test-Verified Defensive Layer for the Model Context Protocol

The Model Context Protocol (MCP) lets AI assistants such as Claude, Cursor and VS Code call external "servers" that can read local files, use the user's credentials and inject instructions into the model's context. Every server a user installs therefore expands their attack surface, yet most users install them with little scrutiny. This report presents GuardiaMCP, a free, open-source and fully offline defensive layer for MCP on Windows. GuardiaMCP combines a static scanner that inspects the MCP servers, skills and configurations already present on a machine against 93 pattern rules, and a real-time "Shield" — a transparent stdio proxy that checks every message between the AI and a local server, hiding tools that change, redacting secrets and blocking poisoned calls, recording every decision in a tamper-evident signed ledger. The tool never executes the code it analyzes, never uses the network, and keeps no account or telemetry. It is developed with a purple-team methodology, and every design guarantee is bound to an automated test. We describe the threat model, the design, the detection approach, and an honest account of what the tool does not cover.