altrusianco/ironcontext

IronContext is a security and optimization tool for AI agents that use the Model Context Protocol (MCP). MCP is a standard way for AI assistants to discover and use external tools. Each tool comes with a description written in natural language that gets loaded directly into the AI's working memory, its context window. This creates a security risk: a maliciously crafted tool description can act like a hidden instruction, tricking the AI into doing something harmful, a technique called prompt injection. IronContext scans those tool descriptions before or during use, checking them against a set of known attack patterns (the May 2026 CVE pattern pack, rules CC-001 through CC-010) to catch things like invisible Unicode characters, script mixing tricks that disguise malicious tool names, hidden instructions, and attempts to get the AI to echo secrets. It also assigns each tool a Reasoning Impact Score (RIS) from 0 to 100, estimating how likely a tool's description is to cause the AI to behave unexpectedly or hallucinate. Beyond security, IronContext trims verbose tool descriptions to reduce how many tokens they consume in the AI's context, which lowers cost and can improve reasoning quality. The tool is built in Rust for speed, with scans completing in under 10 milliseconds on a 100 tool manifest. It can be called from a command line, integrated through Python or TypeScript wrappers, or added as a one step GitHub Action for automated CI checks. Results can be output as plain text, JSON, or SARIF format for use with GitHub Code Scanning. The project is Apache 2.0 licensed.