illiahaidar/mcptrustchecker

MCP Trust Checker is a security scanning tool for MCP servers, which are plugins that give AI assistants like Claude access to outside tools and data. The problem it addresses is that an AI reads a tool's description directly, which makes that description a good place to hide malicious instructions, and a single tool that can both read files and make network requests is already enough to quietly leak your data. This scanner reads the actual published source code of an MCP server, not just its self reported description, and produces a letter grade from A to F along with a numeric trust score out of 100. Its central method looks at each tool in a server and asks what role it could play in an attack: does it bring in untrusted content, does it have access to sensitive data, or can it send information somewhere external. It then builds a map of how tools connect to each other and looks for the dangerous combination of all three roles being present at once, whether inside one tool or spread across several tools working together. It also decodes hidden text that attackers can smuggle inside unicode characters, turning an invisible instruction into something you can actually read and evaluate. The scanner can check tools you already have installed with zero setup, or fetch and inspect the exact published package from npm or PyPI before you install it, verifying it against the official registry hash first. It remembers the hash of a server it has already approved, so if that same version is ever silently swapped out with different code later, a rescan will flag it immediately. Everything runs offline on your own machine: there is no account, no external API calls, and no AI model involved in the scoring itself, so the same server always produces the exact same score. It ships as a command line tool you can run directly, an embeddable library other tools or marketplaces can build on, and outputs results in the SARIF format that many continuous integration systems understand, letting teams fail a build automatically when a server's score falls below a threshold they choose. The project is released under the MIT license and includes hundreds of automated tests plus a written methodology document describing exactly how scores are calculated.