Most browser automation sends a screenshot to a vision model on every step, which is slow and costly. This skill builds an index of the page's controls instead, and Jev chooses which one to use. Only words already in your goal can be typed.
It replays tests, runs goals, generates test cases and explores sites, treating what is visibly on the page as the result. Part of CodexQA, a set of local agent skills that check whether code actually works after it is written.
Instructions in English and Simplified Chinese, with an honest page of known limitations.
What you get
- Full TypeScript source code, MIT licensed
- Agent skill with how-it-works and limitations docs
Jev
TypeScript
Node.js
Claude Code
Codex
CodeRabbit participant
Next.js
Hono
MCP