No account. No API key. No data leaving your machine. Gemma Gem is an open-source Chrome extension that loads Google's Gemma 4 2B model directly in your browser using WebGPU - the browser's hardware-accelerated graphics pipeline, which modern browsers can also use to run AI computations locally.
Once the model loads, a small chat overlay appears on every webpage you visit. The extension gives the model tools to interact with whatever page it's sitting on: read content, take screenshots, click elements, type text, scroll, and run JavaScript. It's a local AI assistant that can actually do things on a page, not just answer questions about it.
What It Handles Well
Ask it to summarize the main argument on a webpage - it reads the page and responds. Ask it to click a signup button or fill a form field - it does it. A thinking mode shows the model's chain-of-thought reasoning as it works, which helps you understand what it's doing and diagnose when it goes wrong.
The developer is upfront about the limits: it's a 2B parameter model. Parameter count is a rough measure of model size and capability - GPT-4 class models are estimated at around 1.7 trillion parameters, making Gemma 4 2B a much smaller system running on local hardware. It handles simple, well-defined tasks reliably. Multi-step reasoning, nuanced analysis, and tasks requiring broad world knowledge are outside its reliable range. Page summarization, basic Q&A, and simple form automation work well. A five-step research workflow won't.
The Case for Keeping It Local
Cloud-based AI assistants - ChatGPT, Claude, Perplexity - are dramatically more capable, but every prompt goes to a remote server. Gemma Gem stays on your machine. For anyone reviewing sensitive documents - legal materials, internal reports, patient records - that's a meaningful difference. There's no usage log being generated, no data being processed under a third-party terms of service, no questions about what happens to your queries.
The extension requires a machine with a modern GPU that supports WebGPU. First load requires downloading the model weights, which takes a few minutes. After that, inference (generating a response) runs entirely locally. The project is open source on GitHub, and for developers interested in browser-based AI experimentation, it's a concrete demonstration of where local model inference currently sits - and where the ceiling is.