Gemma Gem - A Chrome Extension That Runs Google Gemma 4 Fully On-Device in the Browser

This is a Chrome extension that runs Google's Gemma 4 model fully on-device in the browser using WebGPU, functioning like a personal AI assistant.

• No API key, no cloud transmission — no data leaves the device.
• It supports agentic behavior, reading the page you're visiting and performing button clicks, form input, scrolling, and even JavaScript execution.
• You can choose between Gemma 4 E2B (about 500MB) and E4B (about 1.5GB) models, which are cached and reused after the first run.
• It supports q4f16 quantization and a 128K context.

There are 6 built-in tools.

• read_page_content: reads page text, HTML, based on CSS selector
• take_screenshot: captures the visible page as PNG
• click_element / type_text / scroll_page: DOM manipulation based on CSS selector
• run_javascript: executes JS in the page context with full DOM access

The architecture consists of 3 layers: Offscreen Document + Service Worker + Content Script.

• The Offscreen document hosts the model and agent loop using @huggingface/transformers + WebGPU.
• The Service worker handles message routing and screenshot/JS execution.
• The Content script injects the gem icon and shadow DOM chat overlay.

The settings support model switching, Thinking toggle, tool call loop limit, conversation history reset, and per-site disabling.

The agent/ directory is separated based on ModelBackend and ToolExecutor interfaces without dependencies, so it can also be used as a standalone library.

The requirement is Chrome with WebGPU support, the license is Apache-2.0, and it's implemented in TypeScript.