Gemma Gem: Your On-Device AI Browser Assistant

Gemma Gem is a personal AI assistant that integrates seamlessly into your browser, leveraging Google's Gemma 4 model to operate entirely on-device via WebGPU. This means no reliance on cloud services or external API keys, ensuring your data remains private. Gemma Gem can read web pages, interact with elements, execute JavaScript, and answer questions about any site you visit. To get started, you need Chrome with WebGPU support and some disk space for the model. Installation involves a few simple steps using pnpm, and once set up, you can access the assistant by clicking the gem icon on any webpage.

The architecture of Gemma Gem is composed of an offscreen document that hosts the model, a service worker for message routing and executing tasks like screenshots and JavaScript, and a content script for the user interface and DOM interactions. This setup allows the assistant to perform a variety of actions such as reading page content, clicking elements, typing text, and scrolling pages.

Gemma Gem offers flexibility with settings that allow you to switch between different model sizes, toggle thinking modes, and manage conversation history. Development tools are robust, with logging options available for debugging and a tech stack that includes WXT for the Chrome extension framework, @huggingface/transformers for ML inference, and marked for markdown rendering.

For developers, Gemma Gem provides extensive logging capabilities, with logs accessible through Chrome's extension inspection tools. The offscreen document logs are particularly useful for monitoring model operations and tool executions. The agent directory is designed to be modular, allowing for easy extraction and customization.