A new Chrome extension demonstrates on-device AI web automation, running entirely in the browser without cloud APIs or API keys. The tool uses WebLLM with WebGPU acceleration for local inference, executing tasks like navigating to YouTube and playing videos. It's built with a multi-agent system, including a planner and navigator, to handle complex web interactions privately.

The project, called Local Browser, is an open-source proof-of-concept inspired by Nanobrowser. It leverages Alibaba's Qwen 2.5 3B model for reasoning and supports vision-based navigation via Transformers.js for screenshot analysis. The extension requires Chrome 124+ and a GPU with WebGPU support, caching the ~1GB model after the initial download for offline use.

This approach addresses growing privacy concerns in AI automation by keeping all processing local. It also sidesteps cloud dependency and associated costs. The developer plans to add Web SDK support soon, building on existing mobile SDKs. The tool signals a shift toward more capable, private client-side AI agents for everyday browser tasks.