Question 1

Is this a fetch with extra steps, or does it really drive the page?

Accepted Answer

It drives a real headless Chromium through Playwright. Jev picks one action per step from the page's clickable, typeable, and selectable elements, and scores how likely it is that the goal is met or the run is stuck. Your code owns the loop budgets, recovery, and stop gates.

Question 2

Will it submit forms or type my passwords for me?

Accepted Answer

It filled a contact form and stopped without submitting it. For authenticated pages you pick one of two routes: a password fill where code writes the field and the agent types nothing, or seed cookies read from a file so the run starts logged in. The agent never types into a password field, and there is deliberately no cookie flag that takes a value in argv.

Question 3

Is that four-second Wikipedia run a real measurement?

Accepted Answer

It is the project's own figure: Coffee to Espresso in about four seconds for $0.0016, alongside pulling a price off a live pricing page and returning a guide page as markdown. The README also calls the project early software with rough edges on harder sites, so treat those runs as demonstrations rather than a performance guarantee.

Question 4

Will installing it pull down a browser, and does it expose anything on my network?

Accepted Answer

Playwright's Chromium downloads automatically during install unless you set JEV_BROWSER_SKIP_BROWSER_DOWNLOAD=1. Stdio is the default transport. HTTP mode binds HOST to 127.0.0.1 and requires JEV_BROWSER_AUTH_TOKEN once you serve beyond loopback, since every call spends your Jev key and drives a real browser.

Question 5

How do I install and run it, and what license is it under?

Accepted Answer

MIT-licensed and published as @jkudish/jev-browser on npm, so an MCP client runs it as `npx -y @jkudish/jev-browser`. The same agent runs from the CLI with `run "" ` and prints result JSON to stdout, or you import `navigate`, which starts no server and no browser until you call it. The npm package also ships an agent skill under skills/jev-browser/.

Question 6

What comes back at the end of a run?

Accepted Answer

The final page, a step trace with per-step confidences, console errors, and a screenshot. The library hands back the status, the final URL, and page content in the format you asked for, markdown included, and the CLI prints the same result as JSON on stdout. You can also record Playwright output with `--record path.webm` or write an image with `--screenshot path.jpg`.

Question 7

What are the limits when I host one server for a team?

Accepted Answer

Each in-flight call runs its own headless Chromium for up to max_seconds, default 180 and capped at 600, so admitted requests are capped at JEV_BROWSER_MAX_CONCURRENCY, default 4, with the excess getting a 429. Give your proxy read timeouts at least as long as your largest max_seconds, raise client tool-call timeouts to match since the TypeScript SDK caps at 60 seconds, and terminate TLS at your own proxy because the server speaks plain HTTP.

Jev Browser — jkudish

About this project

Highlights

Questions about this entry

Reactions & coverage

Other listings of this project

From the Jev guides

From the Jev AI model community