Small decisions. Interesting possibilities.Submit contentSubmit
BrowserPaw preview

BrowserPaw

Control an everyday Chrome browser with typed Jev decisions and explicit safety checks.

Added to Jevfast

How it uses Jev

BrowserPaw sends a task goal, current page URL and title, bounded accessibility-tree elements and recent history to Jev. Choice questions select the next browser action and target; Noul questions check completion, stalled progress and destructive actions. The engine validates confidence and target indices before executing a step or escalating.

What Jev decides

Illustrative next-action and safety questionsExample answers · not a recorded Jev response · Source ↗
Question 1 · action
YOUR APP
INSTRUCTION

Choose the single browser action that advances the task.

STATE

The task goal, page URL and title, up to 250 filtered accessibility-tree lines, the last five history entries and an optional modal marker. Two representative questions from the seven-question request are shown; answers are illustrative.

JEV · CHOICE
  1. click
  2. type
  3. select
  4. scroll_down
  5. scroll_up
  6. back
  7. wait
  8. done
  9. escalate
Question 2 · destructive
YOUR APP
INSTRUCTION

Check whether the proposed action is destructive or irreversible, such as payment, deletion or posting.

STATE

The task goal, page URL and title, up to 250 filtered accessibility-tree lines, the last five history entries and an optional modal marker. Two representative questions from the seven-question request are shown; answers are illustrative.

JEV · NOUL
YesNo

App workflow

  1. Read the current page

    The engine calls chrome_read_dom for the active viewport, with a 100-element limit and fast format, then extracts the page URL, title, and accessibility-tree lines.

  2. Prepare and query Jev

    It builds the bounded page state and seven parallel System One questions, then sends them to the Jev client.

  3. Validate and apply safeguards

    It checks completion and stuck Noul thresholds, escalates when the destructive Noul reaches 0.5, validates the action Choice, and stops when action confidence is below its configured threshold.

  4. Resolve the target and continue

    For click, type, or select, it validates the chosen element index and confidence, escalates for no target or an ambiguous target, and checks pause keywords and destructive target text before proceeding.

Why it is interesting

The browser loop combines semantic action selection with deterministic safeguards. It limits each task, filters sensitive page data and stops for uncertain targets or destructive actions while using the browser session already logged in by the user.