Read the browser state
The agent reads interactive elements from the page’s DOM, preserving their identity and checking that the page has not changed before acting.
Explore a browser agent that turns a natural-language goal into structured actions in Chrome. Jev decides the next step. Code carries it out.
The agent reads interactive elements from the page’s DOM, preserving their identity and checking that the page has not changed before acting.
Jev selects the next action. A separate text model handles generated text, such as a city name to type into a search field.
Python and browser-harness use a Chrome DevTools Protocol session to click, type, select and scroll. The local inspector shows state and action history.
The core agent, original demonstrations and performance measurements are credited to Gregor Žunič and Browser Use’s Jev Ultrafast project. The upstream MIT licence and copyright notice are retained.
This page does not claim original authorship of the core agent or reproduce upstream benchmarks as Ankit’s results.
Read the upstream performance notes ↗Use Python, Chrome and your own TypeSafe and text-model API credentials to explore the local application. Follow the repository’s installation and configuration instructions.