jev-ultrafast

What it does

jev-ultrafast is a browser agent from the Browser Use team. You give it a start URL and a goal in plain English. On every step it reads the page into a numbered table of the controls you can see, and one Jev request picks both the operation (click, type, select, scroll, wait, done, or blocked) and the target element for it. When the operation is typing, a small LLM writes the text. Nothing else goes to a generative model, and no screenshots are sent in the default loop.

The authors report a full Google Flights search (Zurich to London) in about 7 seconds, plus quick runs on a Wikipedia lookup and a local hotel-search page. Their performance notes are careful to say this is a handful of runs on a few sites, not a reliability benchmark.

Setup

git clone https://github.com/browser-use/jev-ultrafast.git
cd jev-ultrafast
uv sync
cp .env.example .env   # add TYPESAFE_API_KEY and TEXT_MODEL_API_KEY
uv run jev             # local inspector at http://127.0.0.1:8766

It needs Python 3.12 and two keys: TypeSafe for decisions, and an OpenAI-compatible text model for typing. The example config uses an OpenRouter key and Inception’s mercury-2.5. Chrome is controlled through Browser Use’s own Browser Harness, pinned to one version, which asks you to allow remote debugging in Chrome.

Tip for a safer setup: Browser Harness honors a BU_CDP_URL environment variable, which the README doesn’t mention. We started a separate headless Chromium with its own throwaway profile and pointed BU_CDP_URL at it, and the agent ran there without touching any real browser.

In our run, all 31 offline tests passed, and the author’s guard script passed all 21 checks against a real browser without any model calls.

Using it

We ran a Wikipedia task, “from the Kurt Gödel article, open the article about his incompleteness theorems”, against the live site with a real Jev key.

• Without a text-model key, the run crashed. The agent decided to type, and the program exited with a Python traceback: “TYPE_TEXT needs TEXT_MODEL_API_KEY”. The message is clear, but a missing key would be friendlier checked at startup, or returned as a blocked status.
• With a text helper, the loop ran end to end. We substituted a stub OpenAI-compatible server that always returns the same text. The agent observed the page, acted on Jev’s choices, typed the stub’s text, navigated, and reported done on the requested article.
• The inspector is worth using. uv run jev opens a local page that shows the numbered elements, Jev’s operation and target choices, and each executed action, with a step-by-step mode. It’s the best way to understand why the agent did what it did.
• The scope is honest. The authors list what’s out of scope for this MVP: shadow roots, frames, canvas, file uploads, pop-up tabs, nested scrolling, and complex keyboard widgets. An open issue also notes that clickable elements that aren’t native controls, like some autocomplete suggestion rows, can be missed.

Permissions and data

• It uses your real browser by default. Browser Harness connects to your running Chrome, and the agent’s tabs share your profile, including cookies and logged-in accounts. A community issue flags the same concern. Until you’ve read the code, point it at a separate browser (see the tip above), and never run it on pages where a wrong click could send, buy, or delete something.
• Page text goes to TypeSafe on every step. Each request includes the goal and the visible text and controls of the page. Off-screen content isn’t sent.
• Typing steps also go to a text model. That request includes the goal, the target field, the page title, up to 6,000 characters of visible page text, and the last few actions. If you set a text-model key but leave TEXT_MODEL_BASE_URL unset, the code defaults to DeepSeek’s API rather than the OpenRouter endpoint in the example config. Set the base URL explicitly so you know where your pages go.
• Model output is never executed as code. Jev can only choose from elements the agent observed. The text helper must return a small JSON object before anything is typed. Before acting, the executor rechecks that the page hasn’t changed and the target isn’t covered, and runs are capped at 60 browser actions.
• The inspector stays local. It listens on loopback only and checks each request’s host, origin, and a local request token.
• Browser Harness has opt-out telemetry. Its command-line tool (for example browser-harness --doctor, which the README suggests) sends anonymous usage events to PostHog and filters out URLs, page text, prompts, and credentials. The parts jev-ultrafast uses directly didn’t call it in the code we read. Set BH_TELEMETRY=0 to turn it off.

Maintenance

jev-ultrafast is an MIT-licensed showcase from the company behind Browser Use, and the top of the README promotes the Browser Use Cloud waitlist. The public history is only a few commits, but there’s a queue of open pull requests and a small number of open issues. Expect it to evolve quickly or be folded into Browser Use’s main products.

Verdict

This is the most technically ambitious Jev project we’ve reviewed so far, and it’s well built. The safety design is thoughtful: model output never becomes selectors or code, and every action is revalidated against the live page. For developers, it’s also a well-documented example of Jev’s speculative fan-out pattern, where one request answers several questions at once. Run it in a separate browser profile, set both keys and the text-model endpoint explicitly, and treat it as a promising MVP rather than a hands-off agent for real accounts.

For a macOS desktop version of the same idea, see typesafe-computer-use. For browser automation you can call from Claude Code over MCP, see Jev Browser. All options are in browser and computer use.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 19, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.