Gemma Gem is a Chrome extension that runs Google's Gemma 4 (2 billion parameter) model entirely on-device via WebGPU, eliminating the need for API keys, cloud services, or data leaving your machine. It embeds a small chat overlay on every webpage, allowing users to query page content, take screenshots, click elements, type text, scroll, and execute JavaScript directly through the browser. While it offers significant local AI capabilities, multi-step tool chains remain unreliable, and the agent loop sometimes ignores its tools entirely.

Technically, Gemma Gem leverages an offscreen document service worker content script architecture. The offscreen document hosts the model via Hugging Face's transformers and WebGPU, while the service worker routes messages between the content script and the model. The content script injects the chat UI and executes DOM tools like reading page content, clicking elements, typing text, and scrolling. Development builds are available via `pnpm build`, with production builds silencing logs.

The extension requires Chrome with WebGPU support (~500MB disk space for the cached model) and can be installed from the Chrome extensions page (developer mode). It demonstrates a viable path for local AI agents within browsers but highlights ongoing challenges in reliably chaining complex tools for multi-step tasks.