Jev Ultrafast
Web browsing agent published by the Browser Use team, MIT licensed, in Python. Instead of screenshotting the page and guessing where to click, it reads a numbered table of the elements that can actually be acted on (buttons, lists, fields with their value) on every turn, and a small decision model picks one of eight operations (click, type, select, scroll up, scroll down, wait, done, blocked) and a number. A text model is only called when something must be written. The repository reports a Zurich-to-London Google Flights search in 7.1 seconds.
Strengths
- Closed action space: the agent picks a number from the page list, it can neither invent a selector nor write code
- Fast because lean: one network round trip per decision, and ten times fewer browser calls than the previous version according to the repository
- Importable as a Python library, with a local inspector to watch every decision
Limitations
- Openly a demo: iframes, canvases, file uploads and new tabs are not handled
- Two API keys to supply, one of them at TypeSafe for the decision model
- The advertised times come from three runs on one profile, not a reliability measurement
Best for
- Understanding how to bound what an agent is allowed to do, before building one
- Automating a simple, repetitive web path on a site with no iframe or file upload