More Jev Browser Agents: jev-browser-use and Its Forks

Naming collisions are a good sign that an idea is obvious enough to be worth building. In the weeks after Jev's release, at least five separate developers built something called jev-browser or jev-browser-use — a tool that hands Jev the repetitive part of browser automation (which element, which action, is the goal done) and keeps a bigger model or a human for the parts that need judgment. None of them are forks of each other; they're independent answers to the same brief.

We covered the highest-profile browser agent, jev-ultrafast, separately. This article rounds up the rest of the family: one Codex-integration tool with a reported 5–10× speedup, and four standalone jev-browser projects ranging from a documented CLI with real numbers down to a 4-star early build. Star counts and figures below are as of September 2026 — check the Awesome Jev radar for current counts.

Key Takeaways

• jev-browser-use (wy-coliney) is a Codex plugin, not a standalone agent. It requires an existing Computer Use connection and reports "~5–10× faster" navigation in the author's own EZCollegeApp workflows.
• jev-browser (jkudish) publishes the most complete numbers of the group: a 4-second, $0.0016 Wikipedia navigation demo, default thresholds, and explicit v0.1 scope limits.
• jev-browser (Ying-Kai-Liao) reports a 40/42 task pass rate — the author's own site-specific test, not a general success-rate guarantee.
• jev-browser (tontoko) adds a shared CLI, MCP server and TypeScript SDK around the same Playwright-plus-Jev pattern.
• jev-browser (Mrlyk) is the thinnest entry here — 4 stars, no independently verified claims beyond its own description.

jev-browser-use: a Codex speed layer, not a browser of its own

jev-browser-use (MIT, 415 stars) doesn't run a browser by itself — it plugs into an existing Codex Computer Use connection to Chrome or Codex's built-in browser, and lets Jev handle the repetitive navigation (clicking, scrolling, toggling) that would otherwise round-trip to Codex on every step. Jev works from accessibility-tree data, not screenshots, and Codex resumes control for text entry and final verification.

The author reports "~5–10× faster browser operations" from their own EZCollegeApp workflows, attributed to avoiding repeated context-switches back to Codex. On cost, the README states Jev input runs about $0.042 per million tokens with free output, against $2–10 per million for capable Codex models — a claimed 48–238× difference for routine navigation logic specifically, not overall task cost. Both figures are the author's own measurements and haven't been independently retested. The README is upfront that page text gets sent to an external provider (users must authorize that), that it's tested mainly on macOS, and that it isn't suited to tasks needing constant visual judgment. Aiskill.market entry: /skills/jev-browser-use-wy-coliney.

jev-browser (jkudish): a documented CLI with a real worked example

jev-browser by jkudish (245 stars) is a standalone Node.js/Playwright agent, available as an MCP server, CLI, or JS library, that takes a task and a URL and drives a headless Chromium browser toward it. Each step makes "one primary Jev call with three questions": which action to take from the current page's clickable and typeable elements, how likely the goal is met, and how likely the run is stuck. A separate, user-selectable model (OpenAI, OpenRouter, Anthropic or Google) handles any actual text generation — Jev itself never writes.

The README's worked example is a Wikipedia navigation completed in 4 seconds for $0.0016, with typical Jev calls costing "a fraction of a cent." Default thresholds are 0.85 for both goal-probability and stuck-probability, with a budget of 24 steps or 180 seconds per run, and up to 240 interactive elements analyzed per step (truncated beyond that). Password fields are filled only by code, never typed by the model. The project explicitly scopes itself as v0.1 and out: iframes, shadow DOM, hover menus and keyboard-only controls are stated as unsupported, and the README calls it "early software" that should "expect rough edges on harder sites." That candor is worth more than the 4-second demo. Aiskill.market entry: /skills/jev-browser-jkudish.

jev-browser (Ying-Kai-Liao): a 40/42 result, clearly labeled as the author's own

jev-browser by Ying-Kai-Liao (MIT, 76 stars) is a browser library, CLI and MCP server where you supply the goal and any text to type, and Jev chooses elements, actions and values, and separately assesses completion, errors and irreversible steps. The project's dataset entry reports 40 of 42 tasks passing at roughly 300ms per Jev call in the author's own site-specific testing — a real number, but one drawn from a test suite the author designed and ran, not a shared or repeated benchmark. Token-cost comparisons in the README are described as estimates rather than guarantees. Aiskill.market entry: /skills/jev-browser-ying-kai-liao.

jev-browser (tontoko) and jev-browser (Mrlyk): smaller, earlier builds

jev-browser by tontoko (Apache-2.0, 7 stars) wraps the same Jev-picks-the-action pattern in a shared CLI, MCP server and TypeScript SDK, with persistent sessions and UI readback support — though session persistence in the UI doesn't by itself prove database persistence, and that distinction matters if you're relying on it across restarts. Aiskill.market entry: /skills/jev-browser-tontoko.

jev-browser by Mrlyk (Apache-2.0, 4 stars) describes itself as a "browser automation CLI for AI agents, powered by the Jev model's millisecond decisions and near-zero inference costs." At 4 stars it's the earliest-stage project in this cluster; its own listing carries no independently verified claims beyond the description, and no benchmark or worked example was available to check against. Treat it as a name to watch, not yet a tool to rely on. Aiskill.market entry: /skills/jev-browser-mrlyk.

What to do next

1. Match the tool to your stack, not the name. "jev-browser" gets you five different architectures — check whether it needs Codex, Playwright alone, or nothing but an API key before adopting one.
2. Re-run any published numbers on your own site. A 40/42 pass rate or a 4-second demo is a starting point, not a warranty; see how to evaluate Jev projects honestly for a checklist.
3. Prefer the projects that state their scope limits. jkudish's explicit v0.1 boundaries (no iframes, no shadow DOM) are more useful than an unqualified feature list.
4. If you're already inside Codex Computer Use, jev-browser-use is the lower-lift option since it plugs into your existing connection rather than standing up a new browser session.
5. Track the field for consolidation. Five independent implementations of the same idea within weeks usually converges — watch Awesome Jev and our wiring guide if you'd rather add Jev to an agent you already run than pick a browser-specific tool.