browser-use/jev-ultrafast

Project overview

It addresses slow, inefficient agent decision-making over changing page elements and browser state with an action space indexed to the current page, and includes a local inspector showing numbered elements, operation and target probabilities, and executed actions with an option to pause before execution.

Project type

Workflow

Deployment

Refer to project documentation

License

MIT

Best for

• Developers and AI engineers who want a browser agent that adapts its available actions to the observed page state for natural-language tasks.
• Users who want to inspect agent behavior locally, including element numbering, action probabilities, and executed steps, with the option to pause before execution.

Key capabilities

• Generates an element table from each observation and offers compatible CLICK, TYPE_TEXT, SELECT, SCROLL, WAIT, DONE, and BLOCKED operations for browser execution.
• Shows numbered elements, operation probabilities, target probabilities, and executed actions, with an option to pause before execution.

Limitations and risks

• The DOM reader handles common HTML and ARIA controls, not the full accessible-name specification.
• Shadow roots, frames, canvas, uploads, pop-up tabs, nested scrolling, and arbitrary keyboard widgets remain outside this MVP.
• The reported benchmark is three repeats of one task on one browser profile, not a general reliability benchmark.
• The documented flight example checks the route, date, and results but does not select or book a flight.
• The documented demo depends on external TypeSafe and text-model API keys and configured model endpoints.
• A DONE choice still requires independent outcome verification.
• Owned tabs share the existing Chrome profile.

Getting started

• Clone the repository, cd into it, run uv sync, copy .env.example to .env, add TYPESAFE_API_KEY and TEXT_MODEL_API_KEY, run uv run jev, open http://127.0.0.1:8766, then click Start demo → Run automatically. A separate browser harness doctor command and possible Chrome remote-debugging permission may be needed.

Evidence and sources

• README: A browser agent with a dynamic, indexed action space.
• README: The operations are `CLICK`, `TYPE_TEXT`, `SELECT`, `SCROLL_UP`, `SCROLL_DOWN`, `WAIT`, `DONE`, and `BLOCKED`. Only supported operations and targets are offered.
• README: A `DONE` choice still requires independent outcome verification. The DOM reader handles common HTML and ARIA controls, not the full accessible-name specification. Shadow roots, fr…
• README: This is three repeats of one task on one browser profile, not a general reliability benchmark.
• README: It does not select or book a flight.

Source updated at: 2026-09-25T05:00:23Z

Open original GitHub project