Jev for Chrome
A browser extension uses typed decisions to choose actions from the current page.
Added to Jevfast
How it uses Jev
The agent builds state from the task, current page URL/title/text (text capped at 6,000 characters), observed elements, up to ten recent actions, and optional run metadata. Jev receives a Choice question named `operation`, Noul questions `goal_done` and `stuck`, and operation-specific Choice target questions such as `click_target`, `type_text_target`, and `select_target` when those target groups exist. Operation choices are built from current targets and controls, plus DONE and BLOCKED.
What Jev decides
Choose a browser operation and check task progressExample answers · not a recorded Jev response · Source ↗
Question 1 · operation
YOUR APP
INSTRUCTION
Choose one operation supported by the current page, or DONE when all task requirements are visibly satisfied, or BLOCKED when no supported operation can progress. For this representative snapshot, these are the available operation keys.
STATE
Representative source-derived request shape for a hypothetical page snapshot with one click target (element 1), one editable field (element 2), and one dropdown option (target 3:1), and no additional controls. State contains task, page URL/title/text, observed elements, recent actions, and optional run metadata. Dynamic targets and operations are generated from that snapshot; this is not a captured page or Jev response.
JEV · CHOICE
- CLICK
- TYPE_TEXT
- SELECT
- DONE
- BLOCKED
Question 2 · goal done
YOUR APP
INSTRUCTION
Is the entire task already achieved on the current page? Require visible evidence for every requirement; a filled but unsubmitted form does not count as done.
STATE
Representative source-derived request shape for a hypothetical page snapshot with one click target (element 1), one editable field (element 2), and one dropdown option (target 3:1), and no additional controls. State contains task, page URL/title/text, observed elements, recent actions, and optional run metadata. Dynamic targets and operations are generated from that snapshot; this is not a captured page or Jev response.
JEV · NOUL
YesNo
Question 3 · stuck
YOUR APP
INSTRUCTION
Are recent actions failing to advance the task, such as repeating controls or returning to the same page state? The first two steps do not count as stuck.
STATE
Representative source-derived request shape for a hypothetical page snapshot with one click target (element 1), one editable field (element 2), and one dropdown option (target 3:1), and no additional controls. State contains task, page URL/title/text, observed elements, recent actions, and optional run metadata. Dynamic targets and operations are generated from that snapshot; this is not a captured page or Jev response.
JEV · NOUL
YesNo
App workflow
- Observe the task and current page
The builder makes compact state from the task, page URL/title/text, observed elements, recent action history, and optional run metadata. Page text is capped at 6,000 characters.
- Build operation and target questions
The builder maps observed click, fill, and select actions to CLICK, TYPE_TEXT, and SELECT target groups. It adds available controls from the snapshot and always offers DONE and BLOCKED. Each target question is keyed by operation and contains only that operation’s observed targets.
- Check the decision and resolve its target
The agent calls the configured provider, validates the operation, and reads goal_done/stuck. Contradicted terminal choices are vetoed. For a nonterminal operation it validates the matching operation-specific target before proceeding.
- Supply text, act, and observe again
TYPE_TEXT uses a helper to derive the field value from the goal and current context. The agent executes the selected action, records it in history, and observes the page again.
Why it is interesting
The operation list is grounded in controls actually observed on the page, and each target question contains only valid targets for that operation. Jev's DONE and BLOCKED selections are also checked against separate Noul answers before the agent ends the run. This couples a flexible browser action space to explicit progress checks.