← ClaudeAtlas

jev-browserlisted

Drive a real browser with Jev, TypeSafe's ~100ms decision model. Give it a URL and a goal; Jev clicks, types and selects its way there while you watch, then returns the final page. Use when the user asks to use Jev, to browse or navigate a site "with Jev", or wants a fast visible browser demo (search a site, fill a form, open a result).
zurfyx/jev-browser-skill · ★ 3 · AI & Automation · score 76
Install: claude install-skill zurfyx/jev-browser-skill
# Jev Browser `scripts/jev.mjs` opens Chrome, observes the page, and asks Jev (TypeSafe's single-forward-pass model) to choose the next operation and target. It loops until Jev answers `DONE` or `BLOCKED`. You write the goal; Jev makes every navigation decision. If the user asks how it works, point them at README.md and `scripts/core.mjs`, which is the whole decision logic in 250 lines. ## Run it ```bash node <this-skill-dir>/scripts/jev.mjs \ --url "https://en.wikipedia.org" \ --goal "Search for Alan Turing and open his article. Done when the article is visible." \ --text "Alan Turing" ``` Requirements: Node 22+, any Chromium-based browser, and `TYPESAFE_API_KEY` (environment, or a `.env` file in this skill's directory). If the key is missing, ask the user for one — keys come from https://typesafe.ai — and do not try to work around it. ## Writing a good call - **`--goal`**: one or two plain sentences covering the whole task, ending with what "done" looks like on screen ("Done when search results are visible"). - **`--text`**: Jev chooses, it never generates. Pass every string the task needs typed, one `--text` each: search queries, names, form values. Extract them from the user's request. Jev decides which value goes in which field. With no `--text`, typing is not offered. - **`--url`**: start as close to the task as you can (the site itself, not a search engine). - `--screenshot out.png` saves the final page; `--headless` hides the window; `--close` closes