Question 1

How is this different from a screenshot-based computer-use agent?

Accepted Answer

There is no vision model. Each step Jev sees the page as url, text and elements, and picks from a legal action list (click, type, select, scroll, go back, navigate, wait, stop) that code filters before the request is made. Targets are element ids and typed values come from inputs you supply, so the model selects among candidates rather than composing text.

Question 2

Can it type my password without leaking it to the model?

Accepted Answer

Yes, via --secret. The value is typed into the page but never included in the request and never written to disk. Anything else that must be typed belongs in --input, with quoted strings in the goal added automatically. The recorded demo that signs in and then fills a dropdown exists to show the password reaching the page as a typed value rather than model output.

Question 3

What stops it looping forever or clicking something irreversible?

Accepted Answer

Code, not the model. Legal actions are filtered before the request, tried and blocked edges are remembered, loops are detected, and step, cost and time budgets are enforced. A non-none blocker at 0.6 hands the browser back to you for a login, CAPTCHA or consent dialog, and the final page is verified before success is reported rather than assumed.

Question 4

How much does a run cost, and what has to be installed?

Accepted Answer

Roughly $0.0003 a step at about a second, so the recorded three-step run cost a fraction of a cent and the multi-hop Wikipedia run took six steps. You need Node 22+, a TYPESAFE_API_KEY and a browser backend: ego lite by default for signed-in sessions, or --backend chrome for a dedicated headless instance with a JSON result and screenshot. There are zero npm dependencies.

Question 5

Where does install put things, and can I undo it?

Accepted Answer

The installer targets claude-code, codex, agents on the skills.sh convention, cursor and claude-desktop by default, each symlinked into the right skills directory, with cursor-mcp and codex-mcp off unless you name them. It is idempotent, previews with --dry-run, backs up every file it edits and reverts with --uninstall. Other routes are the Claude Code plugin marketplace, npx skills add for skills.sh, or copying skills/jev-browser/ anywhere your agent reads skills from; on Windows prefer install --copy since symlinks need Developer Mode.

Question 6

What commands are there besides run?

Accepted Answer

observe prints the page as Jev sees it, optionally with a screenshot. judge is a raw System One call taking --state or --state-file and --questions or --questions-file, with no browser involved. pick runs one Choice over named candidates with --question, repeated --candidate id=desc, --context and --no-none. doctor is the environment check, with --offline to skip the live API probe and --json for machine output. config shows, prints the path, sets or unsets keys and can take the key from the environment. mcp serves over stdio.

Question 7

What do the demo runs actually show, and can I reproduce them?

Accepted Answer

Three recorded runs, each a single jev-browser run command made with --step-screenshots. A multi-hop Wikipedia task took 6 steps, $0.0029 and 21 seconds. Signing in with a secret and then filling a form with a dropdown took 10 steps, $0.0013 and 18 seconds. Choosing between three identical Start free trial buttons took 3 steps, $0.0003 and 3 seconds, with Jev picking the Team plan from page structure alone at probability 1.00. Any of them can be re-rendered with node scripts/make-demo.mjs.

Jev Browser Skill — ChenYCL

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community