MechInterp Studio
An end-to-end mechanistic interpretability platform. Train sparse autoencoders, discover what your model's features mean, and prove it with causal steering — all from one database-backed workbench, all on your own hardware.
🧠 New to interpretability?
Superposition, features, SAEs, and why steering proves causation — the ideas behind the tool, without the math.
🔬 The Core Workflow
The six-stage pipeline: model → dataset → activations → SAE training → feature discovery → steering.
⚙️ Six SAE Frameworks
Standard, Anthropic, JumpReLU, TopK, Skip, and Transcoder — paper-grounded defaults and tuning guides.
🏷️ LLM-Powered Labeling
Bulk auto-labeling plus two-pass enhanced labeling, using local LLMs or OpenAI.
🌐 Neuronpedia & SAELens
Export findings as complete packages or push directly to a local Neuronpedia instance.
📖 Reference
The full REST API, every WebSocket channel, and the database schema — everything the UI does, scriptable.
What is miStudio?
miStudio replaces the fragmented tooling typically associated with AI-safety research — Jupyter notebooks, custom scripts, and manually tracked experiments — with a professional workbench. Every dataset, training run, feature, label, and steering experiment is a database record with full provenance, and every heavy job runs in the background with live progress.
It runs on your own hardware (a single consumer GPU is enough to start), deploys via Docker Compose or Kubernetes, and never sends your data anywhere unless you explicitly share it.
Finding your way
| If you want to… | Go to |
|---|---|
| Run the whole pipeline once, fast | Quickstart Tutorial |
| Understand the ideas first | Interpretability Primer |
| Install on your hardware | Installation — or the AI-assisted install guides |
| Learn a specific panel | The Core Workflow section, one page per stage |
| Use pre-trained SAEs (Gemma Scope) | SAE Management |
| Script against the backend | API Reference |
| Fix something | Troubleshooting · FAQ |