Zero setup and fully private: paste a trace and get the exact failing step, loop type and tokens wasted in under a second, all in your browser. Deterministic rules instead of an LLM; works with LangGraph, OpenAI and OpenTelemetry traces, and is benchmarked on 148 human-labelled agent failures. Also available: "Ask the dragon," an optional AI second opinion for Claude users. It reviews only the flagged steps, explains the root cause, and catches reasoning mistakes the rules can't see.
