Asked straight, is Jev open source? No — Jev itself is not open source and you cannot run the real model locally; it is TypeSafe AI's hosted System One model, reached only through their API — but the ecosystem around Jev is aggressively open source, and OpenJev now gives you an open approximation of the pattern that you can run anywhere, including on local models. That one sentence settles about ninety percent of the arguments happening this week, so let us unpack the remaining ten.
📺 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
Below: exactly what TypeSafe has and has not published, what running Jev locally actually means in practice, the five open projects worth your attention, and how OpenJev rebuilds the decision loop without a single call to TypeSafe's servers.
Is Jev Open Source?
No, Jev is not open source. TypeSafe AI has not published the weights, the parameter count, or any of the architecture internals. There is no repository that contains the model, no checkpoint to download, and no licence file to squint at — because nothing of the model itself has been released.
What TypeSafe has disclosed is behavioural, not structural. Early access opened on 15 September 2026 through a waitlist, with the console at console.typesafe.ai. The company, founded by Diogo Almeida, has said four concrete things about the hosted model: it was trained with RLCD — Reinforcement Learning for Calibrated Decisions — it evaluates your declared answer set in parallel instead of generating tokens one at a time, it returns decisions in 70 to 500 milliseconds end to end, and it costs $0.042 per million input tokens with output free.
That is the whole public picture. Size, training data, how the calibration works under the hood — all of it stays behind the hosted API. I keep a running file on what is actually documented about the model, and a broader primer in my Jev AI overview, but the verdict does not move: closed model, hosted only, reachable only through TypeSafe.
If you are deciding what this means for your own agent stack, that is the exact conversation running daily inside AI Profit Boardroom — 3,000+ members testing Jev, the open source glue around it, and everything in between, the same week it ships.
Can You Run Jev Locally?
No — you cannot run the real Jev model locally, because there are no weights to download. Every Jev decision is an API round trip to TypeSafe's hosted endpoint. That is not a temporary gap while they prepare a release; it is the product design.
Here is the nuance most people miss: every "local Jev setup" you have seen is already mostly local. The glue is yours. The voice browser runs on your machine. The MCP servers run on your machine. Your workflow code, your browser session, your action loop — all local. The only thing that leaves your computer is the decision itself: a declared set of options goes up, a calibrated choice comes back, usually in well under a second.
So the honest answer splits three ways. The real hosted model: no, never local — the weights only exist inside TypeSafe. The glue around it: always local, in every single setup. And the pattern itself — fixed answer spaces, parallel scoring, calibrated picks — yes, that can now run entirely on your machine through OpenJev pointed at a locally served model. Details two sections down.
📺 Watch: Orca: FREE AI Coding Agent (Open Source!)
The Open Source Shelf Around Jev
While the Jev model stays closed, the tooling around it went open within days of launch. Five projects matter right now — I walked the full shelf in my Jev GitHub ecosystem tour, but here is the short version:
- jev-voice-browser by Moritz Kremb, MIT licensed — voice-controls a real Chromium browser by pairing Jev decisions with Playwright.
- jev-browser by jkudish, also MIT — browser automation as an MCP server, with Jev picking one action per step.
- jev-ultrafast — the official Browser Use project: a dynamic indexed action space, two decisions packed into one round trip, and a documented 7.1-second Google Flights run.
- awesome-jev by yibie — the curated list tracking the whole scene as it lands.
- langchain-typesafe — the official LangChain integration package, for wiring Jev decisions into chains you already run.
Notice what all five have in common: they are glue. Excellent glue, MIT licensed where the repos say so, free to fork — but every one of them still phones TypeSafe's hosted endpoint for the actual decision. Open ecosystem, closed core.
📺 Watch: North Mini Code is INSANE (FREE + Local + Open Source)!
OpenJev: The Open Approximation You Can Run Anywhere
This is the project that changes the local answer. SiliconLabAI/OpenJev landed its first commit on 20 September 2026, and the README describes it as an "open approximation of TypeSafe Jev — a System One style decision engine".
OpenJev rebuilds the contract, not the model: fixed answer spaces you declare up front, per-option parallel scoring, logit-to-softmax probability distributions, and the familiar choice, score, and noul response shape — all reconstructed on top of ordinary chat models.
Setup is three commands' worth of effort: npm install, then npm run dev, then open http://localhost:3001 in your browser. And here is the line that matters for the local question: the model field accepts anything — gpt-4o-mini, qwen-3.8-27b, whatever you prefer — against any base URL you choose. Point that base URL at a model served on your own machine and the entire decision loop runs locally. No TypeSafe account, no waitlist, no data leaving the room.
The README is refreshingly honest about the gap, and the honesty is worth keeping front and centre:
This is "not real Jev" — ordinary LLMs acting as micro-scorers, an architectural approximation of the contract, not a weight-compatible reimplementation.
The difference is real Jev's RLCD calibration and its 70 to 500 millisecond latency. Those live in TypeSafe's weights, and no amount of clever scaffolding reproduces them. What OpenJev hands you is the shape of the thing — same declared options, same scored distributions, same one-decision-per-call discipline — which is exactly what you want for prototyping. I published a full walkthrough in my OpenJev deep dive.
Two cautions before you install. First, read the code — OpenJev is a brand-new repo, and any young project that sits between you and your model endpoints deserves a proper look before it touches your setup. Second, I have not benchmarked it yet; when I do, the verdict lands on Goldie Bench alongside everything else I test.
Open vs Hosted: The Side-by-Side
| Question | Hosted Jev from TypeSafe | OpenJev plus a local model |
|---|---|---|
| Open source? | No — no weights, parameter count, or internals published | Code is public on GitHub; README shows no licence yet |
| Runs offline? | No — every decision is an API round trip | Yes, with a locally served model behind your base URL |
| Calibration | RLCD-trained for calibrated decisions | Ordinary chat models acting as micro-scorers |
| Latency | 70–500ms end to end | Depends on your model and hardware |
| Cost | $0.042 per million input tokens, output free | Your own compute, or your provider's usual rates |
| Access | Waitlist since 15 September 2026, console at console.typesafe.ai | Clone and install — no waitlist, no account |
What To Pick When
Prototyping, learning the pattern, or working under strict privacy rules: go OpenJev with a local model. You get the full decision-loop discipline — declared answer spaces, scored options, one decision per call — with zero waitlist and zero bytes leaving your machine. It is also the cheapest classroom available for understanding why System One decision models feel nothing like chat models.
Production, where latency and decision quality decide whether the agent ships: hosted Jev. The RLCD calibration and the sub-second round trips are the product, and at $0.042 per million input tokens with output free, cost will not be your objection. For where a decision engine actually slots into an agent stack, my Agent OS guide maps the whole architecture.
Not sure which side of that line your use case falls on? Book a free strategy session and we will map it against what is already working for other builders.
FAQ: Jev, Open Source, and Local Setups
Will TypeSafe Open-Source Jev?
Nothing has been announced. TypeSafe has published no weights, no parameter count, no architecture internals, and no statement about opening any of it later. Anyone promising you an open release is guessing — until TypeSafe says otherwise, the hosted API is the only real Jev there is.
What Licence Is OpenJev Under?
The OpenJev README does not state a licence at the time of writing. That matters: jev-voice-browser and jev-browser both display MIT clearly, but OpenJev's repo shows nothing yet. Check the repository yourself before building anything serious on top of it — until a licence appears, you have nothing in writing.
Is the Jev API Key Free?
Access runs through the early-access waitlist that opened on 15 September 2026, managed from console.typesafe.ai. Once you are in, input costs $0.042 per million tokens and output is free — near-zero cost in practice, but still a gated hosted service rather than a free-for-all. I broke the whole process down in how to use Jev AI for free.
The Bottom Line on Jev and Open Source
Is Jev open source? The model, no. The ecosystem, very. Can you run Jev locally? The real model, never — there are no weights to run. But the glue was always local, and OpenJev pointed at a local model now runs the entire pattern offline, end to end.
If you want to be in the room where this gets tested the week it drops — builds, benchmarks, and the occasional faceplant included — join us inside AI Profit Boardroom at $69/mo locked in (normally $110). Or if you would rather talk through your specific Jev play first, grab a free strategy session and we will build the plan together.
