Jev Ultrafast GitHub: Every Feature + Honest Verdict

Jev Ultrafast GitHub points to one official repo, github.com/browser-use/jev-ultrafast, and my verdict is that it is a genuine, MIT-licensed browser agent from Browser Use that is impressively fast on paper but still an early 0.1.0 build.

It is not a faster Jev model, and nobody has put Jev's weights on GitHub.

That is the verdict in two lines.

The rest of this page is the reference version.

I'll go through what it is, who makes it, every feature in the repo, the speed numbers and who measured them, the costs, the limits, and a scores table at the end.

I'll also tell you exactly what I checked myself and what I'm taking from Browser Use's own report.

The Verdict At A Glance

What Jev Ultrafast Is

Jev Ultrafast is a program that drives a real Chrome browser from one plain-English goal.

It reads the page, lists the buttons and boxes it can see, and asks Jev which one to act on.

Jev is a decision model from TypeSafe AI that launched on 15 September 2026.

Jev doesn't write text, because its whole job is to pick an answer from a list and say how sure it is.

A browser agent turns out to be a very good fit for that skill.

Clicking a button is just choosing one item from a list of items on the page.

When the agent does need words, such as a city name for a search box, it hands that one small job to a cheap writing model.

Who Makes It, And Is The Jev Ultrafast GitHub Repo Legit?

Yes, the repo is legitimate.

I pulled these details straight from GitHub on 11 October 2026.

Browser Use is the team behind the browser-use project, which has well over 100,000 GitHub stars, and Browser Harness.

So this is an established team shipping a new idea, and it is not an anonymous upload.

TypeSafe is not the publisher.

TypeSafe supplies the model through its API, and Browser Use wrote the agent around it.

🔥 Want my exact Jev Ultrafast GitHub setup, step by step? Inside the AI Profit Boardroom you get step-by-step video trainings, my installable Agent OS and four coaching calls a week with 3,400+ members building real automations. → Join the AI Profit Boardroom

Feature By Feature: What Is In The Repo

This is the full reference list, based on the README and the code.

One Goal In, No Site Scripts

You give the agent a starting address and a goal in normal language.

The README says there are no site-specific action scripts and no prepared field text in the policy.

That means the flight demo was not a hard-coded recording of a flight search.

A Fresh Numbered Element Table

Every time the agent looks at the page, it builds a new numbered table of the controls it can see.

Each row has a number, a type such as button or textbox, a name, and the current value.

Jev chooses from those numbers, so it never has to describe where something is on screen.

Eight Operations

The agent offers Jev eight operations, and it only offers the ones that make sense on the current page.

Two Decisions In One Request

This is the feature that gives the project its speed.

In a single call, Jev is asked which operation to use and which target each operation would use.

The agent then uses only the target that matches the chosen operation.

The README calls the extra target questions speculative, and it links to TypeSafe's own notes on that pattern.

A Small Text Model For Typing Only

A separate model writes text only when the operation is TYPE_TEXT.

The example settings use Mercury 2.5 through OpenRouter with reasoning switched off.

The README says Gemini, GLM and DeepSeek models can be used through the same OpenAI-style helper.

No Screenshots In The Default Loop

The agent sends Jev structured text about the page instead of images.

Only visible text is sent, so long article bodies and footers that are off screen don't fill the request.

Guards Before Every Click

The agent rechecks that the page is still fresh and that the target isn't covered by something else before it acts.

The README also states that model output never becomes selectors, coordinates, shell commands or JavaScript.

I like this design a lot, because it removes a whole class of things that could go wrong.

A Local Inspector

Running the demo opens a local page at http://127.0.0.1:8766.

It shows the numbered elements, the probability Jev gave each operation, the probability for each target, and the actions taken.

A Choose next button pauses before each action so you can step through a run.

A Python Library

You can import the Agent class and loop over its progress in your own script.

The repo also ships a general runner that takes any address and any goal from the command line.

Small Enough To Read

The README lists six core files, covering the loop, the page reader, the browser connection, the model calls, the instructions and the demo.

For a tool that controls your browser, being able to read all of it in one sitting is a real feature.

Offline Tests

The test suite runs offline and makes no paid calls.

There is also a guard-check script that tests real controls in a local browser without calling a model.

Speed: The Published Numbers

I need to be exact about the source here.

I have not re-run these benchmarks myself, so every number below is from Browser Use's performance report in the repo.

The report also compares the original loop with the optimised one.

That is a 25% lower middle time, and the number of browser calls dropped from 1,092 to 101.

Now for the fine print, which Browser Use wrote itself.

The timer starts after the first look at the page, so browser setup and the first page load are not counted.

Three pairs of runs is a small sample, and the report says so directly.

It also says this is not a broad agent benchmark.

My reading is that the speed is real for this task, and unproven for yours until you test it.

🔥 Want to see how I test tools like this before I trust them? Inside the AI Profit Boardroom, I share my agent builds and test notes as I go, plus weekly coaching calls where 3,400+ members compare what is actually working. → Get access here

Cost

The flights recording used 90,558 TypeSafe input tokens.

By my own maths that is a fraction of a cent at the listed price, but the report does not give a billed dollar amount for the Jev side.

You need your own TypeSafe key, because the code calls TypeSafe's address directly.

That means the free OpenCode Zen model from my free Jev API key roundup will not plug in without a code change.

My post on the Jev waitlist covers getting access in the first place.

Setup: How Hard Is It?

Setup is six commands for someone who already uses Python.

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

You then open the local inspector and click Start demo, followed by Run automatically.

Chrome connects through Browser Harness, and you have to allow remote debugging when Chrome asks.

If the connection fails, uv run browser-harness --doctor is the troubleshooting command from the README.

For a non-technical person this is not a one-click install.

Limits And Weak Spots

The README has a section called "Evidence and limits", and it is refreshingly blunt.

• A DONE choice still needs independent checking, because the agent saying it is finished is not proof.
• The page reader handles common HTML and ARIA controls, and not every unusual component.
• Shadow roots, frames, canvas, uploads, pop-up tabs, nested scrolling and custom keyboard widgets are not supported in this version.
• The tabs it opens share your existing Chrome profile.
• A valid operation can still be the wrong one.
• Live examples and recording scripts make paid API calls.

The shared Chrome profile is my biggest concern for normal users.

An agent that can reach every site you are logged into deserves a separate, clean profile.

The second concern is dependence on one hosted model.

If TypeSafe changes its price or access, the official version changes with it.

The Cloud Version And The Waitlist

The top of the README links to a Browser Use Cloud waitlist for ultrafast browser agents.

On 11 October 2026 that page was headed "SUPERFAST mode" and said it is coming soon to Browser Use Cloud.

It shows a recorded demo and a short form.

There is no price and no date on it, so I'm leaving both out of this verdict.

Alternatives And Lookalikes On GitHub

Searching GitHub for the name brings back more than 50 results.

Here is how I'd sort them.

The Laya port is the interesting one for people who want local decisions.

Its README says it only runs on Apple Silicon and that it had to add extra rules, because Laya is weaker at the open question of what to do next.

If local models are your goal, my local Jev verdict compares Jev, Laya and Clef as decision models.

My Scores

These scores are my judgement from reading the repo, the code and the published measurements.

They are not from a hands-on benchmark of my own, so treat them as a paper verdict.

Who Should Use It

Use it if you are a developer or operator who wants to study the fastest browser-agent design I have seen written down.

Use it if your target task is search, filter and navigate on ordinary web pages.

Wait if you need uploads, pop-ups or complex web apps.

Wait for the cloud version if you don't want to touch a terminal.

And if you like reviews in this format, my DeepSeek Harness review uses the same approach.

Related Reading

• Free Jev API key: the free tiers, tested
• Local Jev verdict: Laya vs Clef vs Jev
• The Jev waitlist explained

📺 Video notes + links to the tools 👉

🎥 Learn how I make these videos 👉

🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉

🔥 Not sure which agent tools are worth your time? Bring your shortlist to one of the four weekly coaching calls inside the AI Profit Boardroom, where 3,400+ members are testing these tools on real work. → Join the AI Profit Boardroom

Also On Our Network

• 🌐 Set up Jev Ultrafast step by step
• 🌐 Seven business uses for Jev Ultrafast
• 🌐 A quick start for Jev Ultrafast in an Agent OS
• 🌐 The models behind Jev Ultrafast, with leaderboard numbers
• 🌐 My original Jev AI guide

FAQ: Jev Ultrafast GitHub Questions

Is the Jev Ultrafast GitHub repo official?

Yes, github.com/browser-use/jev-ultrafast is the official repo and it is owned by Browser Use.

Other repos with the same name are forks or community ports.

Is Jev Ultrafast open source?

Yes, the agent is open source under the MIT licence.

The Jev model it calls is not open source, because TypeSafe has not released its weights.

Is Jev Ultrafast made by TypeSafe?

No, Browser Use made it.

TypeSafe makes the Jev model that the agent calls for each decision.

How fast is Jev Ultrafast?

Browser Use published a 7.073 second Google Flights search, a 2.798 second Wikipedia task and a 1.896 second local hotel search.

Those are the team's own measurements on a small number of runs.

What can't Jev Ultrafast do yet?

It doesn't support shadow roots, frames, canvas, file uploads, pop-up tabs, nested scrolling or custom keyboard widgets in this version.

It also can't prove a task is complete, so you need your own check.

Is there a hosted version?

No, there is no hosted version yet.

Browser Use has a waitlist page that says a fast mode is coming soon to Browser Use Cloud, with no price or date listed.

About Julian

I'm Julian Goldie, an AI entrepreneur, SEO expert, and founder of the AI Profit Boardroom, which has 3,400+ members.

I help business owners scale with AI agents, automation, and SEO.

• I have 400,000+ YouTube subscribers who watch my AI tool tests every week.
• I built a 7-figure agency from the ground up.
• I run daily AI training inside the Boardroom.
• I wrote two Amazon best-sellers on SEO and agency growth.

→ Get my best AI training inside the AI Profit Boardroom

Final verdict: the code is real, free and well designed, but it is an early build with clear gaps.

Test it in a clean Chrome profile on one simple task before you rely on it, and only ever clone the official Jev Ultrafast GitHub repo.