MCPAudit

Security scores and tool-definition history for Model Context Protocol servers.

An MCP client injects every tool's name, description and JSON schema straight into the model's context. A tool description is therefore untrusted input arriving with system authority. MCPAudit reads what each server tells the model, checks it against eight attack classes, and — the part a local scanner structurally cannot do — fingerprints every tool on every scan, so a server that ships clean descriptions, gets adopted, then quietly rewrites them gets caught.