Gemma Gem embeds Google's Gemma 4 model directly in Chrome via WebGPU, enabling page reading, form filling, and JS execution with zero data leaving the device.
An indie developer released Gemma Gem, an open-source Chrome extension that runs Google's Gemma 4 model (E2B ~500MB or E4B ~1.5GB) entirely on-device using WebGPU via Hugging Face Transformers.js. The extension can read pages, click buttons, fill forms, and execute JavaScript — acting as a local browser agent. No API keys, no backend, no data leaves the machine. It's available now on GitHub with a pnpm build workflow and Chrome developer mode install.
This is a working reference architecture for on-device browser agents: WebGPU inference via Transformers.js, an offscreen document hosting the model loop, and a service worker routing tool calls. The agent/ directory is explicitly dependency-free and interface-based — it's designed to be extracted and reused. This is not a demo, it's a buildable pattern for privacy-first browser tooling.
Clone the repo, run pnpm build, load it as an unpacked Chrome extension, and test the JS execution tool against a real form-heavy site — measure token throughput on E2B vs E4B to determine which model fits your use case.
Clone: git clone https://github.com/search for 'Gemma Gem' on GitHub, then run: pnpm install && pnpm build
Tags
Sources
Signals by role
Also today
Tools mentioned