Gemma Gem: On-Device AI Assistant

Chrome extension running Google's Gemma 4 locally in the browser via WebGPU. No cloud services, no API keys.

Capabilities

• Reads and analyzes webpage content
• Clicks buttons, interacts with page elements
• Fills forms, executes JavaScript
• Answers questions about websites
• All processing on-device, no data transmission

Models

• Gemma 4 E2B (~500MB storage)
• Gemma 4 E4B (~1.5GB storage)

Architecture

1. Offscreen Document -- hosts model via HuggingFace Transformers + WebGPU, manages agent loop and token streaming
2. Service Worker -- routes messages, handles screenshots and JS execution
3. Content Script -- chat interface and DOM tools

Tools

Read page content (CSS selectors), capture screenshots, click elements, type text, scroll, execute JavaScript.

Requirements

• E2B: 4GB GPU VRAM, 6-8GB system RAM
• E4B: 6GB GPU VRAM, 8-16GB system RAM
• Chrome 113+ or Edge 113+ with WebGPU

Built with WXT (Vite), HuggingFace Transformers.js, marked.