Gregor Žunič writes a minimal browser-agent framework built around a dynamic, indexed action space and a single TypeSafe request per decision cycle.
The agent reads structured DOM state rather than screenshots, pairs one network round trip with speculative operation and target heads, and invokes a small LLM only for `TYPE_TEXT`.
The flagship Google Flights example completes a natural-language search in roughly seven seconds with a median 25% reduction in task time and an order-of-magnitude drop in browser protocol calls.
- The repo is primarily Python (~58%) with supporting HTML, JavaScript, and CSS.
- It connects to Chrome through Browser Harness and uses a Browser Use Cloud waitlist for hosted deployments.
- Model output never becomes raw selectors, coordinates, shell commands, or executable JavaScript.
- The text helper supports OpenRouter-compatible models such as Gemini, GLM, and DeepSeek.
- The MVP still lacks support for shadow roots, frames, canvas, uploads, pop-up tabs, and arbitrary keyboard widgets.
The agent reads structured DOM state rather than screenshots, pairs one network round trip with speculative operation and target heads, and invokes a small LLM only for `TYPE_TEXT`.
The flagship Google Flights example completes a natural-language search in roughly seven seconds with a median 25% reduction in task time and an order-of-magnitude drop in browser protocol calls.
- The repo is primarily Python (~58%) with supporting HTML, JavaScript, and CSS.
- It connects to Chrome through Browser Harness and uses a Browser Use Cloud waitlist for hosted deployments.
- Model output never becomes raw selectors, coordinates, shell commands, or executable JavaScript.
- The text helper supports OpenRouter-compatible models such as Gemini, GLM, and DeepSeek.
- The MVP still lacks support for shadow roots, frames, canvas, uploads, pop-up tabs, and arbitrary keyboard widgets.
