Jev Ultrafast Replaces Per-Click LLM Planning With Indexed DOM Actions

Browser Use released Jev Ultrafast, a minimal, MIT-licensed browser agent built around TypeSafe’s Jev decision model. It converts each live page into an indexed set of available actions, then asks Jev to choose an operation and target. The repository’s headline Google Flights demo completes in about 7 seconds at normal playback and reports a cost of roughly $0.0039 per run.

The architecture keeps a general-purpose text model out of most control decisions. Jev handles clicks, selections, scrolling, waiting, and task completion. A small language model generates text only when the chosen operation is TYPE_TEXT.

Jev Ultrafast automating a Google Flights search
The project’s demo searches Google Flights using indexed page controls.

A Fresh Action Table at Every Step

At every step, an atomic DOM snapshot captures visible controls, labels, values, and text in one browser call. Atomic collection means the fields describe the same page state. The agent turns that snapshot into a numbered table such as:

[1] button    Change ticket type - Round trip
[2] combobox  Where from?        - San Francisco
[3] combobox  Where to?          - empty
[4] textbox   Departure          - empty

Jev’s action vocabulary is limited to CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT, DONE, and BLOCKED. Each target index refers to a control present in the current snapshot, and the runtime retains a reference to the corresponding DOM node.