An automated AI red-team. It attacks an AI agent, proves each break with a ground-truth check, fixes it with a guardrail, and re-tests on attacks it never saw.
The offline sample bot is free and instant. Real models use a planted fake secret + a small attack budget (a few cents). Nothing you enter is stored.
Source + 2-minute CLI guide: TEAM.md on the julian/devtool-mvp branch.
julian/devtool-mvp