Drop-in input/output safety for AI agents — screen prompt-injection on the way in, catch leaked secrets, refusals and slop on the way out. Zero dependencies.
A small, dependency-free guard you wrap around your agent's I/O: it inspects prompts for injection attempts before they reach the model, and scans outputs for leaked secrets, refusals, and low-quality 'slop' before they reach the user. Local, fast, MIT-licensed.
Free and open source (MIT). Full install and usage in the README on GitHub.