Community · E2 · artifact verified
Drive a browser with typed page judgments
Give jev-browser a task and a URL: Jev picks one action per step from the page's clickable, typeable, and selectable elements and scores goal-met and stuck likelihood, while code owns budgets, recovery, and stop gates.
01 · Role in the system
What Jev does here
On every step the harness flattens the live page into its interactable elements, and one request asks Jev to choose the next action from them plus two calibrated probabilities - is the goal met, is the run stuck. Jev never improvises a plan string; it answers a Choice and two Nouls, and the trace records each with its confidence. Code owns the loop: step budgets, recovery moves, and the decision to stop. The README reports real-site runs, including Wikipedia Coffee to Espresso in about four seconds for $0.0016. Free text, when an action needs it, is typed by a separate small provider model, keeping Jev to typed judgments.
02 · Control boundary
Where Jev sits
Per step the harness turns the page into candidate actions and asks one Choice plus goal and stuck Nouls; deterministic code validates the pick, executes it, and decides when to stop.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- The author calls it early software with rough edges on harder sites; the costed examples are short navigations, not long tasks.
- Typing goes through a second provider model, so the pipeline is Jev for decisions plus a text model for fields only.
- Action quality depends on the page's own accessibility of elements; pages that hide interactivity degrade the action space.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- jkudish ↗Community · github · public · checked 2026-09-20