jev-ultrafast
A browser agent where Jev selects an operation and a compatible page element from a live indexed action space.
What Jev decides
Which supported browser operation to run, and which compatible visible element to target
Each observation becomes a structured element table. Jev chooses the next browser operation and its compatible target in one TypeSafe request; a small LLM is used only when text must be written.
- Jev role
- Select the next browser operation and its target element
- Input
- The natural-language goal plus an indexed table of visible page controls, names, values, and text
- Output
- One operation and the corresponding compatible target element
- Decision frequency
- Once per browser decision cycle
- Fallback
- A small LLM generates a field value only after Jev selects TYPE_TEXT
Decision example
Illustrative indexed action space
IllustrativeInput
Goal: find a one-way flight from Zürich to London. Visible controls: trip type button, origin combobox, destination combobox, departure textbox.
How it works
Reported performance
Recorded task time
7.073 s
Google Flights: Zürich to London, with generated text and loading waits included
Author reported · Source reviewedCaptured Sep 21, 2026
Optimized median
7.092 s
Three verified runs of one Google Flights task on one existing Chrome profile
Author reported · Source reviewedCaptured Sep 21, 2026
Median browser protocol calls
101 calls
Optimized median, compared with 1,092 calls in the original runtime
Author reported · Source reviewedCaptured Sep 21, 2026
Overview
jev-ultrafast is an open-source Python browser agent from Browser Use. Instead of asking a model to generate browser instructions, it turns the current page into an indexed table of visible controls and asks Jev to choose from the operations and targets that are valid for that state.
Jev selects an operation and a compatible element in one TypeSafe request. When the selected operation is TYPE_TEXT, a separate small language model generates the field value. The default loop uses structured page state rather than screenshots, and the executor rechecks page freshness, target geometry, and click occlusion before acting.
The public performance figures are author-reported and source-reviewed here, not independently reproduced by WhatCanJevDo. The featured Google Flights recording completes in 7.073 seconds. A matched comparison reports a 7.092-second optimized median versus 9.450 seconds for the original runtime, but the authors explicitly describe it as three repeats of one task on one browser profile—not a broad reliability benchmark.
The current MVP handles common HTML and ARIA controls. The project documentation lists shadow roots, frames, canvas, uploads, pop-up tabs, nested scrolling, and arbitrary keyboard widgets as unsupported. A DONE choice still needs independent outcome verification.
- Language
- Python
- License
- MIT
- Author
- Browser Use
Sources
jev-ultrafast MIT License
Browser Use · Author reported · Source reviewed
Faster on the real web — performance report
Browser Use · Author reported · Source reviewed
jev-ultrafast Python package metadata
Browser Use · Author reported · Source reviewed
jev-ultrafast README
Browser Use · Author reported · Source reviewed
Related projects
Cua Driver jev-use
A public-preview Cua Driver recipe for bounded Jev browser actions with independent verification.
What Jev decides
Which supplied action candidate to take next, or whether to reobserve or abstain.
Jev Review
A code-review pipeline assembled from small typed judgments instead of one long reviewer prompt.
What Jev decides
What is risky, which files and evidence matter, what could fail, how severe it is, and who should review it.
Jev Voice Browser
A browser that starts resolving voice commands while the user is still speaking.
What Jev decides
What the user intends, which page target it applies to, whether the command is complete, and whether it is destructive.