Back
browseragentsOpen source

jev-browse

Drive a browser one Jev decision at a time through agent-browser.

What Jev decides

Click, type one supplied value, press Enter, go back, finish, or report blocked.

Jev chooses clicks, supplied text values, navigation actions, and when to stop. agent-browser performs each action.

CHOICENOUL
Jev role
Pick the next browser action and determine whether to finish or report a block.
Input
The user goal, current page content, available actions, and supplied text values.
Output
Typed action and goal-state decisions with probabilities.
Decision frequency
Once per browser step.
Fallback
Stop when the goal is blocked; never invent text or a destination URL.
Author reported · Source reviewed

How it works

Input
The user goal, current page content, available actions, and supplied text values.
Jev
CHOICE + NOUL
Output
Typed action and goal-state decisions with probabilities.

Overview

jev-browse is a CLI and MCP server for goal-directed browser use. Each step asks Jev whether the current page meets the goal, whether the task is blocked, and which available action makes progress. agent-browser executes the selected action. Jev does not generate text or enter arbitrary URLs.

The README reports a 10-task, three-run comparison with Claude Code and provides a rerunnable evaluation script. Its results are author-reported and do not establish performance on longer tasks. The project documents limits around duplicate element labels, CAPTCHAs, dragging, uploads, and downloads.

Language
JavaScript
License
MIT
Author
kyrylosyzonenko
Sources checked Sep 23, 2026

Sources

Related projects