User repository on GitHub
Let Jev, TypeSafe's ~100ms decision model, drive your browser. A plug-and-play skill for Claude Code and Codex.
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Demos & Experiments
Jev Browser Skill — zurfyx
Three dependency-free files that put Jev in a browser agent's hot path: code builds the choices from a table of visible controls, Jev picks an operation and target in one forward pass, and the code executes the pick.
- Stars
- 2
- Language
- JavaScript
- License
- MIT
- Published by
- Community
- Author
- zurfyx
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Jev Browser Skill — hqman.
About this project
Let Jev drive your browser. A plug-and-play skill for Claude Code and Codex: you name a site and a goal, Jev clicks, types and selects its way there, a few hundred milliseconds per decision. Hacker News at 1× speed, one goal: log in, search for "Jev", open the first result's comments. Nine decisions, eight seconds. Watch the MP4 · Step through it.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- JavaScript
Highlights
- Speculative fan-out: one request asks Jev for the operation plus a target for every operation, so two dependent decisions cost a single round trip.
- Jev sees only a numbered table of visible controls from an in-page SNAPSHOT call, never the DOM, and any answer outside the offered set throws.
- A question caps at 255 options, so the skill offers the first 250 controls and warns in the log; narrowing to a region needs the fuller jev-ultrafast.
- Measured on a MacBook with a warm window, a Selenium form with two fields, a dropdown and a checkbox took five steps and 5.1s, 1.4s of it inside Jev.
Watch out
This is deliberately a reference implementation built to be read, not a finished browser agent, and the README says the production-grade version with a live inspector, region narrowing, stale-page guards and a text-generating helper model is the separate browser-use/jev-ultrafast repository that this one follows. There is no model here that writes text: every string you type has to be supplied up front with --text, which is a real limit for any task that needs new prose. A question cannot exceed 255 options, so pages bigger than that get a logged warning rather than full coverage. Install needs Node 22+, a Chromium browser and a TypeSafe API key, and there is no npm install, no Playwright and no browser download. Timings come from a MacBook with a warm window and the rest of each run is the page loading. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
No. A --secret is typed only into password fields and never enters a request; Jev is told a password field exists and whether it is filled, never the value. Ordinary strings arrive as --text values and go into the request as text_values, and Jev picks which value belongs in the chosen field. What Jev cannot do in this repository is write new field text, which the README flags as the difference from jev-ultrafast, where a small helper model handles that.
One request carries several named questions at once, and the code reads only the answer matching the chosen operation. After Jev returns TYPE, the click_target answer is discarded and only type_target is read, so a decision and its target arrive in one round trip instead of two. In the recorded Hacker News trace the eight steps took about 300 to 530 ms each, and end to end the Selenium form was 5.1s with 1.4s inside the model. A step is 3 to 6k input tokens with no charge for output.
You get a truncated choice set and a warning. A question tops out at 255 options, so the skill offers the first 250 controls and says in the log when it skipped some. The Hacker News front page fits with 227 controls; a product grid does not, and needs a first question that narrows to a region before the target question, which is exactly what the linked jev-ultrafast repository adds. By default the skill drives your own browser when it allows remote debugging and otherwise opens a separate Chrome profile that stays open so logins persist.
That was a labeling problem, not a model failure. While building the login demo Jev returned BLOCKED at 0.31 right after the password step. The fields were named acct and pw, and the link and the button were both called login, so the table gave it nothing to separate them. Better labels in the element table, plus roles in the action history, took the same situation to CLICK login at 0.89. Jev is stateless, so the history the code chooses to send is the only memory it has, and JEV_DEBUG=1 prints the full distribution at every step.
A SNAPSHOT call inside the page returns a numbered table of visible controls, and Jev sees only that table plus visible text, never the DOM. The operation question offers only what is possible now: CLICK, TYPE, ENTER, DONE, BLOCKED, WAIT and SELECT, with ENTER appearing only after a TYPE, SELECT absent unless there is a dropdown, and WAIT always on offer. Each target question then offers only the elements that operation makes sense on.
A step is 1k to 30k input tokens, almost all of it the page: a blank page costs 1k and the Hacker News front page 18k, with output free, which the README puts at under a tenth of a cent per decision and about half a cent for the whole nine-step demo. Measured on a MacBook with a fresh window per run: Wikipedia search and open was 2 steps in 3.1s with 0.7s in Jev, the Selenium form 5 steps in 4.2s with 1.0s in Jev, and the Hacker News run 8 steps in 8.1s with 2.2s in Jev.
Node 22+, a Chromium browser such as Chrome, Edge, Brave or Chromium, and a TypeSafe API key, with no npm install, no Playwright and no browser download. By default it drives your own browser when remote debugging is allowed, otherwise it opens a separate Jev window with its own persistent profile so logins survive between runs. To use your own Chrome, enable remote debugging once at chrome://inspect and approve a per-run prompt.
More Jev resources like this
3GitHub stars
Browser agent where Jev makes every navigation decision and agent-browser performs clicks and typing, stopping when the named goal page is reached or blocked.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
kyrylosyzonenko/jev-browseJavaScript
kyrylosyzonenko/jev-browseJavaScript5GitHub stars
An MCP server that lets Claude or Codex drive a real Chrome window: Claude plans the task and Jev picks each click, keystroke and scroll from a numbered element list, with irreversible actions held for confirmation.
Jev Demos & Experiments#community#awesomejev#catalog-import
4GitHub stars
Browser-use layer for CLI coding agents: Jev picks the operation and target element in one round trip per step, exposed over MCP with a CLI, a Python API and an HTTP server.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
brnyxx/jev-raPython
brnyxx/jev-raPython10GitHub stars
Playwright wrapper that routes natural-language operations and assertions through Jev while clicks, typing and waits stay deterministic, with verification provenance.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
tontoko/jev-browserJavaScript
tontoko/jev-browserJavaScript2GitHub stars
Ten runnable JavaScript experiments on agent decisions, each pairing a small Jev judgment with visible application policy code, plus authored fixtures so the whole set runs offline.
Jev Demos & Experiments#community#jevbest#catalog-import
0GitHub stars
A browser MCP server where Jev picks the next page action from locally ranked candidates in roughly 300 ms without LLM tokens, deterministic state diffs verify the step, and the driving agent steps in only on escalation.
Jev Demos & Experiments#community#jevbest#catalog-import
From the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.