# Drowse

> 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.

The browser workbench runs supported models locally with WebGPU. Available instruments depend on the selected model and compatible packs. Python also provides a library, CLI, local HTTP APIs, SAE training and lens fitting.

## Documentation

- [Overview](https://drowse.ai/index.md): Product summary and entry points.
- [Explore language models with Drowse](https://drowse.ai/learn.md): Learn to inspect language models, steer activations, run local WebGPU inference, and compare experiments in the Drowse workbench.
- [Activation steering for language models](https://drowse.ai/learn/activation-steering.md): Explore activation steering without changing model weights: concepts, manifolds, lens words, SAE features, neurons and heads, ablation, and probe gates.
- [Inspect tokens and internal representations](https://drowse.ai/learn/interpretability.md): Inspect token probabilities, concept probes, sparse autoencoder features, R- and S-lens layer reads, and the neurons and heads behind a token in Drowse's local LLM workbench.
- [Run a language model locally in your browser](https://drowse.ai/learn/browser-models.md): Run compatible Gemma, Qwen, GPT-2, and Pythia models locally with Drowse. Learn about WebGPU, downloads, device checks, and browser storage.
- [Compare language-model experiments](https://drowse.ai/learn/experiments.md): Compare language-model continuations with Loom branches, token forks, activation interventions, template scoring, and saved experiment recipes.
- [Drowse for developers and browser agents](https://drowse.ai/developers.md): Use Drowse's experimental WebMCP browser tools or the separate Python library and OpenAI-compatible, Ollama-compatible, and native inference APIs.
- [Privacy and your local data](https://drowse.ai/privacy.md): What stays in your browser with Drowse, and what reaches Hugging Face, Neuronpedia, Cloudflare hosting, browser agents, and the contact form.
- [About Drowse](https://drowse.ai/about.md): Drowse is an open-source workbench for local language models, activation steering, and mechanistic interpretability. Find source, licensing, and contributors.

## Agent operation

- [Drowse browser skill](https://drowse.ai/skills/drowse/SKILL.md): Operate the existing WebMCP tools in a supported browser.
- [Source code](https://github.com/stdlogits/drowse): Implementation, installation, issues and license.

## Optional

- [Combined documentation](https://drowse.ai/llms-full.txt): All public guides in one file.

- [Agent resource catalog](https://drowse.ai/.well-known/ard.json): Machine-readable discovery for the browser-agent skill; not a remote inference endpoint.