Give the task a goal
Start with the same fixed trip as the recorded demo. No travel knowledge or JSON editing needed.
Zürich → London Historical recording
One way · September 20, 2026 · 1 adult · Economy
Stop when matching flight options appear. Nothing is booked.
Give Jev eyes
Attach the current recorded screenshot as an image asset—not a DOM or a list of controls.
Trip progress
One request combines three typed questions. Add the goal and screenshot first.
operationChoice- What should happen next? Choose an operation such as click, type, or finish.
targetBox- Where should it act? Locate the control in the screenshot. Used only for click or type.
text_valueChoice- Which supplied city? Zürich or London. Used only for typing; no text is generated.
Questions JSON
Full request · goal, image asset, and recent actions
No image attachedScreenshot input only · no live website
Decision
Input screenshot & target
Action history 0
Inspect an action to revisit its exact input and answer. This does not rewind the task; use Previous to go back.
Applied actions will appear here.
About this example
Observe → Choose → Act → Repeat. Live Jev receives the screenshot, goal, and recent accepted actions—not a DOM or hidden page state. One request chooses an operation, a target box, and a city supplied by the goal. Only answers needed by the chosen operation execute. A transient loading frame uses a labelled local recorded wait, with no inference call.
Replay is scripted and makes no model calls. Both sources follow the fixed path captured in the GIF; unsupported actions do not change the image. Completion is checked against the recording, not a live website. Dates and fares are historical. Nothing is booked.
Original Jev Ultrafast demo ↗