The Jev model itself is not on GitHub — it is TypeSafe AI's hosted API, closed and cloud-only — but the Jev GitHub scene is already very real, led by Moritz Kremb's open-source jev-voice-browser project (MIT licence) that I demo controlling a real Chromium browser by voice, plus official integrations like the langchain-typesafe package from the LangChain team.
📺 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
So if you typed "jev github" hoping to clone the model weights, save yourself the search. That repo does not exist. What does exist is more interesting: a fast-growing pile of open-source glue that wires Jev's decisions into real tools. In this guide I'll show you exactly what is on GitHub, what is not, how the voice browser repo works under the hood, and the one caveat I give everyone before they install anything.
Is Jev Itself Open Source? No — Here's What Actually Lives on GitHub
Let's set the record straight before you waste an afternoon searching GitHub for weights. Jev is a closed, hosted model from TypeSafe AI, the lab founded by Diogo Almeida. It opened for early access on 15 September 2026, and you reach it one of two ways: TypeSafe's own console at console.typesafe.ai, or through Vercel's AI Gateway. There is no download link, no weights file, no self-hosted option.
Quick refresher on why Jev is different, because it explains why the GitHub ecosystem looks the way it does. Jev is a System One, decisions-only model. It never writes free-form text. You hand it a question and a fixed set of options — up to 255 choices — and it picks one in roughly 70 to 500 milliseconds, at $0.042 per million input tokens with output completely free. I've broken the full model down in my Jev AI overview, so here's the one-table version of what is and is not on GitHub:
| What you searched for | Where it lives | Open source? |
|---|---|---|
| The Jev model itself | console.typesafe.ai or Vercel's AI Gateway | No — hosted API only |
| jev-voice-browser | github.com/moritzkremb/jev-voice-browser | Yes — MIT licence |
| langchain-typesafe | LangChain's official integration package | Yes — open package |
That middle row is the reason "jev github" is suddenly a search term. Let's go deep on it.
Jev GitHub: The Voice Browser Repo Everyone Is Talking About
The flagship of the Jev GitHub ecosystem is jev-voice-browser, an open-source project by developer Moritz Kremb, released under the MIT licence. I ran it on camera this week, and I'll be honest: it's the first voice interface that didn't make me want to crawl back to my keyboard. You talk to a real browser, and it moves before you finish the sentence.
Here's the pipeline in plain English. Your browser's built-in speech-to-text turns your words into a live transcript. Jev reads that transcript plus a snapshot of the page and picks the action. Playwright — the same tool developers use to test websites — then drives a real Chromium browser to carry it out. No vision model squinting at screenshots, no twenty-second pauses. You can grab jev-voice-browser on GitHub and inspect every line of that pipeline yourself, which matters for reasons I'll get to shortly.
The Speed Trick That Makes This Repo Special
Most voice agents wait for you to stop talking, then think. This repo does the opposite, and it's the clever bit worth studying even if you never run it. On every new word you speak, the app re-asks Jev what to do and cancels the stale request from the word before. Because Jev answers decision questions in well under a second and charges nothing for output, hammering the API like this is affordable in a way it never was with a chat model.
Each ask bundles roughly ten questions at once — is this a navigation command, which element does it target, is the user still mid-sentence, and so on. In my own runs it was nine questions coming back in about 400 milliseconds, total. That's why the browser starts moving while you're still talking.
How Jev Clicks the Right Thing Without Writing a Word
The repo's other smart pattern: it never lets Jev generate anything. Page snapshots label up to 100 clickable elements, and Jev answers with a label — a pure multiple-choice pick, which is the only thing Jev does. When you dictate text, your own sentence pieces get pointed at and reused, never rewritten, so the model cannot hallucinate words into your search box. And when Jev's confidence on a pick drops under 0.45, the app doesn't guess: it pins numbered badges on the candidate elements and asks you which one you meant.
That is decisions-only architecture done properly, and it's exactly the pattern we build with inside AI Profit Boardroom — 3,000+ members shipping agent systems like this every single week. If you want the agent-side theory first, my Jev AI agent guide covers when a fast decider beats a chat model.
📺 Watch: Gemma 4 DESTROYS Github Copilot?
Installing the Voice Browser: The Four-Line Reality and My One Caveat
Install friction on this repo is genuinely low. From Moritz's instructions it's about four lines in a terminal, and then a local page inside Chrome becomes your control panel for a Chromium window. You'll need API access — a TypeSafe key from console.typesafe.ai — and the project can also run through an OpenRouter key if that's how you already manage your models.
Now the caveat, and I said this in the video because it matters more than the demo:
Read through the code before you install anything. Every repo, every time — not just this one.
That's not paranoia, it's hygiene. jev-voice-browser is MIT-licensed and small enough to skim in twenty minutes, so you can see exactly what it sends and where it sends it. That's the whole point of the Jev ecosystem living on GitHub instead of inside someone's closed app: you get to check. So check.
And if you'd rather have a second pair of eyes on your setup — which repos are safe to build on, where Jev fits your stack, what to automate first — book a free strategy session and I'll map it out with you.
📺 Watch: NEW Perplexity Computer Update is INSANE! (Claude Code + GitHub CLI)
The Official Integration Packages: langchain-typesafe
Community repos are one half of the Jev GitHub story. The other half is official integrations, and the big one so far is langchain-typesafe — LangChain's own package for wiring Jev into chains and agents.
It gives you state management plus Jev's three question types — Choice, Score, and Noul — as first-class components. LangChain also shipped a ready-made model-router piece built on Jev, and the design is refreshingly human: you describe each model's strengths in plain English, and Jev routes every incoming request to whichever model's description fits best. A router that responds in 70 to 500 milliseconds and costs $0.042 per million input tokens is the kind of infrastructure that used to take a fine-tuned classifier and a lost weekend.
Going the API route rather than the voice route? My Jev AI API guide walks through the console, the question types, and the pricing in detail.
What the Jev GitHub Ecosystem Grows Next
I'll frame this honestly: what follows is the pattern, not a promise. Every time a genuinely new model capability ships, GitHub fills up with glue code around it within weeks. With Jev, the glue is decision-layer projects — open-source repos that bolt a fast, cheap decider onto tools you already use. Expect things like:
- Inbox triage that sorts and prioritises mail before you ever open it
- Request routers like LangChain's, picking the right model for each task
- Task boards that reorder themselves as priorities shift
Those aren't hypotheticals — they're the shape of the 10 builds I walk through in my how to use Jev guide, and most are a weekend project for one developer precisely because the model does one narrow thing extremely well.
My own roadmap is on the record too: Agent OS will grow a Jev decider layer as the model matures. The architecture is already agent-first — start with the Agent OS guide if you're new to it — and a decisions-only model slots in as the traffic cop between agents. When that layer lands, I'll test it the way I test everything: against real tasks on Goldie Bench, not vibes.
Want to build alongside people running these exact repos right now? That's what AI Profit Boardroom is for — live builds, the prompts and repos as I ship them, and a room full of members comparing results.
Jev GitHub FAQ
Is Jev open source?
No. Jev is a closed, hosted model from TypeSafe AI. You access it through console.typesafe.ai or Vercel's AI Gateway, and there are no public weights. The open-source part of the Jev GitHub world is the tooling around the model — projects like jev-voice-browser and packages like langchain-typesafe.
Who made the Jev voice browser?
Developer Moritz Kremb built jev-voice-browser and released it on GitHub under the MIT licence. It combines browser speech-to-text, Jev for the decisions, and Playwright driving a real Chromium browser. I demoed it in my video "Jev AI Voice Browser Agent is WILD!" — the repo is exactly what you see on screen.
Can I run Jev locally?
The model itself, no — Jev stays hosted on TypeSafe's infrastructure. The glue runs locally, though: the voice browser runs on your machine, drives your own Chromium, and only the tiny decision questions go out to the API. That split — local open-source glue, hosted decider — is how every project in this ecosystem works today.
How much does the voice browser cost to run?
The repo is free under MIT. The Jev calls behind it cost $0.042 per million input tokens with output free, and the bundled questions are tiny. Even with the re-ask-on-every-word trick, my test sessions cost pennies.
Get the Builds, Not Just the Repo Links
Here's the play. Star the repo, read the code, run the demo — that costs you one evening. The money is in applying the decisions-only pattern to your own business, and that goes a lot faster with help.
Inside AI Profit Boardroom you get my Jev builds as I ship them, the Agent OS decider work as it lands, and 3,000+ members stress-testing the same repos you just read about. It's $69 a month locked in — normally $110 — and your price never rises once you're in.
Prefer a one-on-one look at where Jev fits your stack first? Book a free strategy session and we'll plan it together. Either way, remember the headline: the model is closed, but the ecosystem is wide open. Go read some code.
