OpenInterpretability

About

The inaugural mech interp toolkit operates directly within Claude Code, Cursor, and Cline using MCP. Its production probes, such as FabricationGuard and agent-probe-guard, detect hallucinations and agent malfunctions. The suite includes the ProbeBench leaderboard and offers SAE training, scaling from a free 30-minute Colab notebook to research-level quality. It is released under the Apache-2.0 license.