Jev GitHub: The Voice Browser & Open Builds

Searching for Jev GitHub repos? One honest clarification first: the Jev model itself is not open source — it’s TypeSafe’s hosted API. What IS on GitHub, and growing by the day, is the ecosystem built on top of it — headlined by Moritz Kremb’s MIT-licensed jev-voice-browser, the project that lets you control a real Chromium browser with your voice so fast it moves before you finish the sentence. I’ve run it (that’s it in the video above). Here’s the tour, the install, and the check-the-code rule.

The Jev GitHub project everyone’s installing

The voice browser is three parts wearing one trench coat. Your voice goes into the speech recognition already built into your browser; the words stream to Jev, which makes the decisions; and Playwright — the same tool developers use to test websites — moves a real browser window, clicking and typing like a person. Jev never writes a word here: it answers about ten questions per utterance in one batched call (what do they want? which element? which site? finished speaking? how far to scroll?) — in my runs, nine questions came back in around 400–445 milliseconds.

The design detail that makes it feel telepathic: it re-asks Jev every time a new word arrives, cancelling the previous request — like a waiter who stops writing the moment you change your order. By the time your sentence ends, the answer is usually already waiting. In my small eight-command test it went eight-for-eight at a 445ms median, against 1.6s for Claude Haiku, 2.2s for Gemini 3.8 Flash and 2.5s for GPT-5.6 Luna (seven of eight each) — a small test, and those models weren’t tuned for this, but the gap is the feeling of the thing.

Why the design can’t click what isn’t there

Before every decision the app snapshots the page: up to 100 clickable things, each with a short label like E04 — and Jev’s answer is just one of those labels, so clicking a phantom element is structurally impossible. Same trick for your words: the code slices your sentence into candidate pieces and Jev only points at one, which the code copies letter-for-letter into the search box — a model that can only point at your own words can’t put words in your mouth. Web addresses come from a list the code owns; Jev just picks.

And when it genuinely isn’t sure — confidence under 0.45 — it doesn’t guess: numbered badges appear on the page and you say “two”; that reply doesn’t even go to the model. It’s the same schema-locked philosophy as Jev’s core architecture, applied to a browser.

Install, keys, and the wider Jev GitHub ecosystem

Setup is four lines in a terminal, then a local page inside your Chrome drives Chromium; an OpenRouter key works for access. My standing rule — same one from the 9Router repo guide — applies doubly to young repos that touch keys: read the code before you install. MIT licensing is exactly what makes that possible; use the privilege.

Beyond the headline repo, the ecosystem is sprouting fast: LangChain shipped a ready-made Jev model-router component, Browser Use rebuilt their browser-agent loop around Jev picks, and Alex Volkov published a context-trimming plugin for Claude — the receipts for all of those live in the use-cases breakdown. Access to Jev itself for any of these: the API routes guide.

The bottom line on Jev GitHub

The jev github answer: the model stays closed, the ecosystem is wide open — and Moritz Kremb’s voice browser is the repo that shows what System One speed feels like in your hands: point-don’t-write design, snapshot labels, confidence badges, and a browser that moves before you stop talking. Four lines to install, one read-through first, and the future stops being theoretical.

FAQ: jev github

Is Jev open source on GitHub?

The model isn’t — it’s TypeSafe’s hosted API. The ecosystem on GitHub is, headlined by the MIT-licensed jev-voice-browser.

What is the jev-voice-browser?

Moritz Kremb’s open-source project: browser speech-to-text feeds Jev’s batched decisions, and Playwright executes clicks, typing, scrolling and tabs on a real Chromium window.

Why is it so fast?

Jev returns labels and numbers instead of writing, ~10 questions per batched call, re-queried on every new word with stale requests cancelled — ~400ms per decision in my runs.

Can it hallucinate a click?

It picks from labelled snapshots of what’s actually on the page (max 100 elements), and copies your own words letter-for-letter — phantom clicks and invented text are structurally blocked.

What do I need to run it?

Four terminal lines, Chrome, and a key (OpenRouter works). Read the code first — it’s MIT, so you can.

What else is on GitHub for Jev?

A LangChain router component, Browser Use’s agent loop, and community plugins like Alex Volkov’s context trimmer — all early, all worth auditing.

Related reading

• Jev AI: The Model That Never Writes A Word
• Jev AI API: Every Way To Get Access
• Jev Architecture: How System One Actually Works

Last updated September 2026. This is the living guide to jev github — it gets updated as the tools change.