Project profile
Jev Ultrafast
Browser Use's open-source browser agent. Jev chooses the next operation and an observed target.
Browser UseGitHubDocumentation
Last verified: 2026-09-24
Quick facts
| Project | Jev Ultrafast |
|---|---|
| Creator | Browser Use |
| Type | Browser Agent |
| Language | Python |
| License | MIT |
| Uses Jev for | Operation and target selection |
| Open source | Yes |
| Status | Active |
What it is
Jev Ultrafast turns the visible page into a numbered list of controls, asks hosted Jev for one operation and one target, and uses a separate small text model only when the operation is TYPE_TEXT. The browser harness executes that choice after its own checks.
Why it matters
It is a public example of Jev acting as a fast decision layer inside an agent: the program enumerates legal actions, and Jev picks among them instead of inventing a selector or a script.
How it uses Jev
| Decision role | Pick the next operation and which observed element it applies to. |
|---|---|
| Input state | A fresh numbered table of visible controls, not a screenshot in the default loop. |
| Question type | Choice. Operation and target are one TypeSafe request. |
| Output | An operation such as CLICK, plus an index that was in the snapshot. |
| Fallback model | A small text model writes the string only for TYPE_TEXT. |
| Execution | DOM state, then Jev, then operation and target, then the browser executor. |
Architecture
- DOM state
- Jev
- Operation
- Target
- Browser executor
Key features
- Indexed action space over controls the page actually exposes.
- Targets resolve to nodes from the current snapshot, not to model-written CSS selectors.
- TYPE_TEXT is the only step that calls a text model.
Performance and claims
Browser Use describes the project as a fast, cheap web agent. docs/performance.md times one Google Flights recording at 7.073 seconds. That run searches options and stops; it does not book a flight.
Jev AI Hub has not re-run the recording. The 7.073-second figure and the before/after pairs are the authors' measurements.
Treat 7.073 seconds as one documented demonstration, and the three-pair comparison as a small before/after of their own loop. Neither number is a universal ranking against other browser agents.
Limitations
- The README leaves shadow roots, frames, canvas, uploads, pop-ups, nested scrolling, and arbitrary keyboard widgets outside the MVP.
- A DONE choice still needs independent outcome verification.
- The published timing comparison is one task family, three pairs, with a sign-test p-value of 0.25.
Code and repository
https://github.com/browser-use/jev-ultrafast
GitHub snapshot captured 2026-09-24: 19,100 stars, 1,260 forks, 132 open issues. Last push 2026-09-18. Counts change; this is not a live lookup.
Related Jev concepts
ChoiceSystem One ModelDecision layerBrowser Agents
Related guides
Jev Ultrafast explainedArchitecture and fan-outBrowser-agent tutorial
Related projects
No other published projects share this organization, type, or use case yet.
Sources
- jev-ultrafast READMEBrowser Use · accessed 2026-09-24 · github
- Faster on the real webBrowser Use · accessed 2026-09-24 · github
- browser-use/jev-ultrafast repository metadataGitHub · accessed 2026-09-24 · third-party