Jev

Jev is an experimental browser skill that shows how a language model can drive a real browser step by step. Rather than hiding the mechanics, the site walks through an actual run: what the skill listed from the page, what Jev was asked, what it answered, and what executed.

The core loop is simple. The skill reads the current page and lists everything it could click, type into or select. Jev picks one element and states how confident it is; the skill performs the action, then asks again with the updated page until Jev declares the goal done. Each request bundles the operation and a speculative target, so two decisions travel in one round trip.

The site includes replayable traces such as logging into Hacker News and searching, an Amazon add-to-cart and delete flow, and reading a Yahoo chart without typing. A live mode is also offered, and the underlying logic lives in scripts/core.mjs, the same file the skill runs.

It is aimed at developers and researchers who want to understand or prototype LLM-driven browser automation, not at end users seeking a polished assistant. The documented constraints are useful: Jev sees only the element table, not the DOM or a screenshot, and vague labels like acct and pw can produce low-confidence or blocked answers.

Use the traces to study failure modes and confidence behavior before adapting the approach. Overall, Jev is a transparent, minimal reference implementation for AI browser control.