Are the benchmark numbers real?

They are author-run results on 10 short tasks, 3 runs each, with cleared cookies and agent-browser-only limits. Both setups passed 30/30, with lower median time and about 70 times less cost reported for Jev on these tasks. The readme says they do not show longer or more complex tasks.

Why can't it type whatever it wants?

By design Jev answers with probabilities and never writes text. Typing is limited to texts you pass via --text or the MCP texts map, and login works only with a username and password you supply.

What are the sharp edges on real pages?

Duplicate labels expose only the first element, extract returns matching page lines verbatim and can miss some, background reloads can reset typed values unless network wait stays on, and checkout was tested only up to stopping before Finish.

Is it safe to use my real password?

No. Every request sends page content and your texts, including any password, to TypeSafe. The readme tells you to pass only test or throwaway credentials.

What three questions does jev-browse ask Jev each step?

Is the current page the goal, does the goal need an action the agent cannot do, and which action makes the most progress: a click, typing one of your texts, pressing Enter or going back.

What exit codes does the jev-browse CLI return?

0 when the goal was reached, 1 for no useful action or the step limit, 2 for wrong usage, 3 when the goal needs an impossible action, and 4 when the run crashed.

How do I plug jev-browse into an agent?

It ships a stdio MCP server exposing one browse tool that returns a status of reached, stopped, blocked or error, the final URL, a step log and any extracted lines, and it works with Claude Code, Codex and Hermes.

CommunityJev Demos & Experiments

Jev Browse — kyrylosyzonenko

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.

See highlights

Open github.com

Stars: 3
Language: JavaScript
License: MIT
Published by: Community
Author: kyrylosyzonenko
Last verified: Oct 2026

kyrylosyzonenko/jev-browse

User repository on GitHub

View on GitHub

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is Jev Browse — 0x7067.

About this project

jev-browse drives a real browser with TypeSafe Jev making every decision and agent-browser performing every action. You give it a goal, such as open the Hacker News FAQ. Jev picks each click, chooses which of your texts to type into which box, and decides when the goal is reached or can't be reached. jev-browse is an unofficial project and isn't affiliated with TypeSafe or Vercel.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Evaluation Jev for Tool calling

Used inJev for Software & DevTools

Languages: JavaScript

Highlights

Each step sends one Jev request covering goal-reached, blocked, and next-action choice with probabilities.
Jev never writes text; search queries and form values come only from caller-supplied texts or the calling agent.
CLI and single-tool MCP server accept goal, optional URL, texts, step limit, and verbatim extract options.
Short-task benchmark reports 30/30 passes for both agents with lower median time and cost for Jev plus agent-browser.

Watch out

Unofficial and unaffiliated with TypeSafe or Vercel. It types only supplied texts, cannot enter URLs, solve CAPTCHAs, hover, drag, upload, or download, sees only the first of duplicate labels, and sends page content plus texts including passwords to TypeSafe, so use throwaway credentials. Requires Node.js 22.9 or later, agent-browser, and TYPESAFE_API_KEY. 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.

More Jev resources like this

jev-browser

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.