jev-browse
Browser driver where Jev makes every decision and agent-browser performs every action, benchmarked against Claude Code on 10 tasks.
Overview
jev-browse by kyrylosyzonenko drives a real browser with Jev making every decision and Vercel's agent-browser performing every action. Given a goal such as open the Hacker News FAQ, Jev picks each click, chooses which of the supplied texts to type into which box, and decides when the goal is reached or cannot be reached.
Its benchmark compares it with Claude Code (Sonnet 5) driving agent-browser directly on the same 10 tasks, three runs each, with identical limits: agent-browser commands only, no entering URLs, only the supplied texts, and cleared cookies. Tasks include navigation on Wikipedia, Hacker News, example.com and the TypeSafe docs, two typing goals, two goals the agent must refuse such as logging in without credentials, and a checkout on the Sauce Labs demo shop stopping before Finish. Both passed 30/30; jev-browse had a 4.4 s median and $0.0009 per task versus 9.4 s and $0.0679, about 70 times cheaper. The README notes it does not show performance on longer or more complex tasks. It is unofficial and MIT licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which element should be clicked or which supplied text typed into which box? | — |
| Yes / no | Has the goal been reached, or can it not be reached? | yes · no |
How it works
- Give a goal and supplied texts
- agent-browser snapshots the page
- Jev decides the next action or completion
- agent-browser executes; repeat
Strengths
- 30/30 tasks passed in the benchmark
- About 70x cheaper than Claude Code driving the browser
- Correctly refuses impossible goals
Caveats
- Only 10 relatively simple tasks
- Unofficial project
At a glance
- Type
- Benchmark
- Status
- beta
- Language
- TypeScript
- Runtime
- Node CLI with agent-browser
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Drives a real browser with Jev making every decision and Vercel's agent-browser performing every action, with a benchmark.
hellogumbo/awesome-jev
Frequently asked questions
How did it compare with Claude Code?
Both passed 30/30; jev-browse was faster (4.4 s vs 9.4 s median) and about 70x cheaper.
Can it type arbitrary text?
It chooses among texts you supply for the task.
Does it handle long tasks?
The README says the benchmark does not show performance on longer or more complex tasks.
Related projects
Browser-agent benchmark with five timed events from one prompt, plus a ~200-line DevTools agent where Jev picks each action.
ProjectbetaJavaScriptego-decision-layer: measured, pluggable decision layer for ego lite where Jev picks operation and target in one request.
ProjectbetaTypeScriptPlugs any CDP browser, including engines that never draw a page, into a Jev loop that picks operations and targets from the DOM.
LibrarybetaPythonBrowser automation where Jev picks every step from a sightmap's named actions and reports whether the goal is met.
ProjectbetaGoBrowser agent that acts on one look at the page with one open-model call per action; Jev is an optional chooser backend.
AppprototypePythonOpen Apache-2.0 browser-agent decision model (Qwen3.5-35B-A3B fine-tune) that doubled Jev 1.13's task success in the same agent.
ServiceresearchPython