# Drowse | Local LLM interpretability and activation steering

Open-source, fully local LLM interpretability workbench in your browser. Inspect token predictions and steer a model's internal activity.

Drowse is an open-source workbench for mechanistic interpretability, activation steering, and local inference with language models. Run a compatible model in your browser, read every token through probes, sparse features and two lenses, steer it with concepts, words, features or single neurons and heads, and compare how the reply changes. A separate Python library and server support programmatic experiments.

- [Activation steering](https://drowse.ai/learn/activation-steering.md): Steer with concepts, manifolds, lens words, SAE features, neurons and heads, and probe gates.
- [Inspect language models](https://drowse.ai/learn/interpretability.md): Read each token's alternatives, probes, SAE features, the R and S lenses, and the neurons and heads behind it.
- [Run a model in your browser](https://drowse.ai/learn/browser-models.md): Check WebGPU requirements, model choices, downloads, and local storage.
- [Compare experiments](https://drowse.ai/learn/experiments.md): Use branching conversations, token forks, and restricted-choice scoring.
- [Developers and browser agents](https://drowse.ai/developers.md): Discover WebMCP tools and the separate Python, OpenAI-compatible, and Ollama-compatible APIs.
- [Privacy and local data](https://drowse.ai/privacy.md): Understand what stays on your device and which actions make network requests.

[Open Drowse](https://drowse.ai/app) · [About and source](https://drowse.ai/about.md) · [Privacy and data storage](https://drowse.ai/privacy.md)