Skip to content
← The showcaseComplete free read
Public repository

Choose browser actions from an indexed control table

A browser-agent loop asks Jev to choose an operation and live DOM target while a separate model supplies text only when needed.

Editorial workflow map
  1. Index visible controls
  2. Choose operation and target
  3. Verify the resulting page
View the author’s original post (opens in a new tab)
Our interpretation · not the author’s screenshot or a reproduced result

The inspectable decision point

Each step turns visible page controls into indexed candidates. Jev selects an action and compatible element; text generation is isolated to typing.

Supporting documentation (opens in a new tab)

Why this is reusable

The executor rechecks freshness and occlusion instead of treating a model choice as a valid selector. Outcome verification remains a separate responsibility.

Limits stated by the project

The repository describes a narrow benchmark and excludes several interaction classes, including frames, canvas, uploads, popup tabs, and arbitrary keyboard widgets.

Use the same checks on another demo.

Read the complete free evaluation guide →