The Browser Use Jev GitHub scene is led by two repos right now: browser-use/jev-ultrafast, the official Browser Use project that booked a real Google Flights search from Zürich to London in 7.1 seconds by letting TypeSafe's Jev make two decisions in one network round trip, and jkudish/jev-browser, an MIT-licensed MCP server that drives a headless browser for fractions of a cent per task. Two different builders, two different layers, one shared idea: stop asking a slow reasoning model to make every single click.
📺 Watch: NEV Jev AI Voice Browser Agent is WILD!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I pulled both READMEs and verified them on 21 September 2026, and every number below is attributed to the repo that published it. My "Jev AI: 10 INSANE Use Cases" video from 19 September cites the same Browser Use figures. One thing up front: I have not run either project myself yet. Treat everything here as the maintainers' own claims, and read the code before you install anything. That rule never goes away.
Why browser agents needed a decider
The old way of building a browser agent looked like this: screenshot the page, send everything to a big reasoning model, wait a few seconds, get one click back. Then do it again. Every micro-decision — which button, which field, scroll or wait — cost a full LLM round trip. Chain thirty of those together and you have a slow, expensive agent burning tokens on choices that were never really language problems.
Jev is TypeSafe AI's answer: a System One decisions model that picks from options you define in 70 to 500 milliseconds, attaches a confidence score to every answer, and costs $0.042 per million input tokens with output free, per TypeSafe's published pricing. It does not write prose. It decides. I broke the model down in my Jev AI explainer, but here is the part that matters for browsers: a web page is exactly the shape Jev likes. A finite set of elements. One operation per step. GitHub builders spotted the fit within weeks.
Browser Use Jev GitHub: inside jev-ultrafast
browser-use/jev-ultrafast is the official Browser Use org repo, and its README describes a browser agent with a dynamic, indexed action space. Every observation produces a fresh numbered element table — buttons, comboboxes, textboxes with their current values — and the operations on offer are CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT, DONE and BLOCKED. Only supported operations and real targets are ever offered, so the agent cannot pick an element that is not actually on the page.
The clever bit is speculative targeting. One TypeSafe request answers the operation AND the candidate targets — click target, type target, select target — in the same shot. If the operation comes back CLICK, only the click target can execute. Two decisions, one network round trip. That is where the speed lives.
Jev never writes text in this system. A small writing model steps in only when the operation is TYPE_TEXT. The demo config runs on an OpenRouter key with inception/mercury-2.5 as the current text model, reasoning disabled, and the README notes Gemini, GLM and DeepSeek also work through the OpenAI-compatible helper.
The headline run, straight from the jev-ultrafast README: Zürich to London on Google Flights in 7.1 seconds. One natural-language goal, with real text generation and page-loading waits included. No site-specific action scripts. No prepared field strings. Browser Use says the outcome is independently verified, and the labels on the demo screenshots are added afterwards — they never drive the browser.
This is the kind of build we pull apart every week inside AI Profit Boardroom — 3,000+ members wiring browser agents, MCP servers and Jev-style deciders into real businesses instead of just watching demos.
📺 Watch: New Hermes Browser Use Agents are INSANE!
How to try jev-ultrafast
Setup for jev-ultrafast is short. Clone the repo from GitHub, run uv sync, copy the env example, and add TYPESAFE_API_KEY plus TEXT_MODEL_API_KEY. Then uv run jev and open http://127.0.0.1:8766. If you have not sorted access yet, my Jev API guide covers keys and pricing tiers.
What loads at that local address is the best learning tool in the whole Jev ecosystem: an inspector showing the numbered elements, the operation probabilities, the target probabilities and every executed action. A Choose next mode pauses before execution, so you can watch the agent weigh each element before anything moves. Chrome connects through browser-harness, Browser Use's own repo. And if you would rather skip self-hosting entirely, Browser Use Cloud has a waitlist open for ultrafast browser agents in the cloud.
📺 Watch: New OpenAI Browser Use Agent is WILD!
jev-browser: the MIT-licensed MCP server
jkudish/jev-browser attacks the same problem from the integration side. It is MIT-licensed per its badge, and the README pitch is fast, very cheap browser use with Jev: hand it a task and a URL, and it drives a real headless browser through an MCP server, a CLI, or a library you import.
The division of labour is strict, and it is the right one. Jev picks one action per step from the page's clickable, typeable and selectable elements, and at the same time scores how likely the goal is already met or the run is stuck. The code owns the loop — budgets, recovery, stop gates. At the end you get the final page, a step trace with confidences, console errors and a screenshot.
The repo reports five real-site runs, all attributed here to the project itself:
- Wikipedia's Coffee article to the Espresso page in about 4 seconds, for $0.0016
- A contact form filled in — and the agent stopped without submitting it
- A price pulled off a live pricing page
- A guide page returned as clean markdown
- An accessibility-tree breakdown of a WordPress site
That second run is the one I would show a client. Anyone can demo speed. An agent that halts at the submit button is an agent you can eventually trust near real accounts. And the maintainer is upfront that jev-browser is early software with rough edges on harder sites — honesty that reads as a feature, not a warning.
The paste-into-your-agent install
Requirements from the jev-browser README: Node.js 20 or newer, plus a TypeSafe key from console.typesafe.ai — the keys page specifically. The package ships on npm and runs via npx, and Playwright's Chromium downloads automatically on first run, with an env flag to skip it if you already manage browsers yourself.
Here is the detail I genuinely love: the README ships a paste-into-your-coding-agent install prompt. You paste one paragraph into Claude or whichever agent you use, and the agent registers the MCP server and walks you through key setup itself. That is exactly the pattern I push with AI Profit Boardroom members — ask your agent to set the tool up, do not hand-edit config files at midnight. And if you want a second brain on where a browser agent actually fits your business before you install a thing, book a free strategy session and we will map it together.
The wider shelf on GitHub
Beyond the two lead repos, the Browser Use and Jev corner of GitHub is filling in fast. jev-browser-use is a Chrome extension and MCP bridge carrying Jev browser experiments, and awesome-jev curates public Jev projects and integrations in one place. jev-browser already has an active fork family too, including a hardened, production-oriented edition. Add Moritz Kremb's jev-voice-browser, which I have covered before, and this is a real ecosystem now — I keep the full map current in my Jev GitHub roundup.
What this means for your agent stack
Both repos land on the same architecture, and it is the one worth stealing: a fast decider chooses, deterministic code owns the loop, and confidence scores gate anything irreversible. That is the discipline I bake into my Agent OS guide, and browser agents like these are exactly what I want to put through Goldie Bench next — claims are nice, receipts are better. If you are building agents that act on the open web, start with the Jev agent patterns and layer the browser on top, not the other way round.
And if you would rather build alongside people already running these setups, AI Profit Boardroom is $69/mo locked in (normally $110) — 3,000+ members, weekly calls, and working agent stacks you can copy instead of reverse-engineering READMEs alone at 1am.
Browser Use Jev GitHub FAQ
What is jev-ultrafast?
jev-ultrafast is the official Browser Use repo on GitHub pairing their browser agent stack with TypeSafe's Jev decisions model. Per its README, every page observation becomes a numbered element table, one Jev request picks both the operation and the candidate targets in a single round trip, and a small separate model writes text only when typing is required. Its headline claim: a Zürich-to-London Google Flights search completed in 7.1 seconds from one natural-language instruction.
How much does Jev browser automation cost?
Two datapoints, each from its own source. TypeSafe prices Jev at $0.042 per million input tokens with output free, which is why per-step decisions cost close to nothing. And the jev-browser README reports its Wikipedia run — Coffee to Espresso in about 4 seconds — cost $0.0016 total. Neither number is mine; both are the projects' published figures, so verify them against your own runs before you budget anything.
Can my coding agent install it?
Yes — that is the intended path for jev-browser. The README includes an install prompt you paste straight into Claude or any coding agent, and the agent registers the MCP server and handles key setup for you. You still need Node.js 20 or newer, a TypeSafe key from console.typesafe.ai, and npx available. My standing rule: have the agent read the repo's code first and report what it found before anything executes.
The Browser Use Jev GitHub story is moving weekly — new forks, new inspector features, a cloud waitlist already open. If you want a browser agent earning for you instead of sitting in your bookmarks, grab a free strategy session and I will show you exactly where a 7-second agent fits in a real funnel.
