An end-to-end workbench for training, understanding, and using mechanistic-interpretability sparse autoencoders (SAEs) — now with automated feature clustering, cross-layer circuit discovery, resumable labeling that knows what is left and why, an agent-controllable MCP server, …
An OpenAI-compatible local inference server that attaches pretrained sparse autoencoders to a running model and steers its behavior in real time — now serving whole clusters and discovered circuits, and controllable by AI agents over MCP.
Where the mechanistic AI suite is going — from sparse-autoencoder feature discovery toward a general environment for finding a learned system’s internal variables, testing how they interact, and applying only what survives validation at inference time.