Jev GitHub: What TypeSafe Actually Publishes (and What It Doesn't)

You search jev github expecting one repository — the model, maybe some weights, a README with a quickstart. Instead you get thousands of repositories with "jev" in the name, awesome-lists claiming more than a thousand projects, an organization literally called jev-ai with fourteen repos named after search terms, and another called TypeSafeAI that looks official until you read its bio. Two weeks after launch, finding the real thing takes more effort than it should.

Here is the answer up front: TypeSafe AI's official GitHub organization is github.com/typesafe-ai. It publishes client SDKs, an agent skill, evaluation code and integrations. It does not publish Jev itself — no weights, no training code, no architecture — and the repository many people search for, typesafe-ai/jev, does not exist.

We run jev-ai.org, a playground and API built on the Jev model. For this guide we checked every repository in TypeSafe's organization through GitHub's API on September 30, 2026, and cross-checked which ones TypeSafe itself links from typesafe.ai and docs.typesafe.ai. That cross-check is the only reliable test of what is official. If you are still getting your bearings, start with what Jev AI is.

Last updated September 30, 2026. Repository lists, versions and star counts change quickly; the dates in each section say when we checked.

One disambiguation: in medical results, JEV is the Japanese encephalitis virus, and there are unrelated GitHub projects called "jev" that predate the model. This guide is about Jev, the decision model from TypeSafe AI.

Jev GitHub: The Short Answer

Everything in TypeSafe's Official GitHub Organization

The typesafe-ai organization held 11 public repositories when we checked. Six are about Jev; the other five are forks and infrastructure that have nothing to do with using the model.

The forks are worth a sentence because they fuel speculation. A fork of vLLM or of a diffusion-language-model paper in a company's GitHub tells you what its engineers looked at, not what Jev is built on. TypeSafe has not published Jev's architecture, and outside guesses remain guesses — our explainer on how Jev works separates what is documented from what is not.

What Is Not on GitHub

If you came looking for any of these, they are not there, and nothing official suggests they will be soon:

• Model weights. Jev runs only as a hosted API. TypeSafe's models page says the same weights serve every account and are not fine-tuned or LoRA-adapted per customer.
• Training code or data. TypeSafe describes itself as "primarily a data research lab" that makes all of its training data itself, and it has not released it.
• An architecture paper. The launch post mentions a new architecture, a parallel sampler and the RLCD training method, without technical detail.
• A typesafe-ai/jev repository. Plenty of people search for it; the URL is a 404.
• The Doom demo. TypeSafe promised a walkthrough of its Doom clip, but no Doom code was in the organization when we checked. Community versions exist — see why Jev AI can play Doom.

So there is nothing to download and run offline. If that is what you need, our guide to downloading or running Jev AI locally covers what is and is not possible.

Official vs Unofficial: How to Tell

This is the part most "Jev GitHub" pages skip. Within a week of launch, several organizations appeared whose names look official. Here is what we found when we checked each against GitHub's API.

Four checks settle almost every case in under a minute:

1. Is it linked from typesafe.ai or docs.typesafe.ai? TypeSafe's docs link the SDK sources and the agent skill directly. If the only links to a repo come from other GitHub repos or directories, it is not official.
2. Check the package name. The official packages are typesafe-sdk on PyPI and @typesafe-ai/sdk on npm. Anything else calling itself "the Jev SDK" is someone else's wrapper.
3. Look at the creation date and the owner's bio. An organization created days after launch, linking to a website that is not typesafe.ai, is not TypeSafe.
4. Be careful with your API key. Many community projects ask for your TypeSafe key. Run them locally, read the code first, keep the key in an environment variable, and never paste it into a hosted demo you cannot inspect. A key spends your balance.

None of this makes community projects bad. Many are useful — they are just not TypeSafe's, and they are not maintained on TypeSafe's schedule.

How to Use the Official SDKs

Both SDKs take an API key from the TypeSafe console in the TYPESAFE_API_KEY environment variable and call TypeSafe's System One API. These snippets are copied from the SDK READMEs.

Python:

uv add typesafe-sdk

from typesafe_sdk import Choice, TypeSafeClient

with TypeSafeClient() as client:
    response = client.system_one(
        state={"document": "I was charged twice. Please fix this ASAP."},
        questions={
            "category": Choice(
                instructions="What is this ticket about?",
                criteria={"billing": None, "technical": None, "other": None},
            ),
        },
    )

print(response.choices["category"].choice)

pip install typesafe-sdk works too, according to TypeSafe's SDK docs.

JavaScript and TypeScript (Node.js 20 or newer):

npm install @typesafe-ai/sdk

import { choice, TypeSafeClient } from "@typesafe-ai/sdk";

const client = new TypeSafeClient();
const response = await client.systemOne({
  state: { document: "I was charged twice. Please fix this ASAP." },
  questions: {
    category: choice("What is this ticket about?", {
      billing: null,
      technical: null,
      other: null,
    }),
  },
});

console.log(response.answers.category.choice);

Both SDKs are young — the Python one reached v0.7.2 on September 26, 2026, and each has already had a breaking change to how score criteria are passed. Pin the version in your lockfile.

Calling Jev without a TypeSafe account

The SDKs above talk to TypeSafe's own API. If you would rather not set up a separate account, the Jev AI API on jev-ai.org takes the same kind of request — a state and typed questions — over plain HTTP, with no SDK required:

curl https://jev-ai.org/api/v1/systemone/ \
  -H "Authorization: Bearer $JEV_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jev-1.13",
    "state": "Order #8814 arrived with a cracked screen. Third time this quarter.",
    "questions": {
      "refund_requested": {
        "type": "noul",
        "instructions": "Is the customer asking for a refund?"
      }
    }
  }'

The endpoint, key format and response fields differ from TypeSafe's — keys start with sk-glm5-, the path ends in a slash, and each response reports exactly what the call charged. The full contract is in our API docs. New accounts get five free credits, and you can edit and run the preloaded scenarios in the Jev AI playground before writing any code.

The Agent Skill for Claude Code and Codex

The most-starred official repository is not an SDK but the agent skill. It gives a coding agent the context to write correct Jev integrations — the question types, the patterns, the current docs and cookbooks. TypeSafe documents two installs:

claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai

npx skills add typesafe-ai/skills --skill typesafe-ai

Be clear about what this does. It makes your coding agent better at writing code that calls Jev. It does not turn Jev into the model behind Claude Code or Cursor — TypeSafe's docs state plainly that Jev is not a drop-in replacement for a coding agent's LLM. For the full setup, including routing through a gateway, see our guide to using Jev with Claude Code and Cursor.

Reproducing TypeSafe's Evals from GitHub

WorkflowEvals is the repository behind TypeSafe's published workflow benchmarks. It covers four workflows — security incidents, agent traces, invoices and customer service — and runs Jev or any supported LLM against the same harness:

uv sync --locked
export TYPESAFE_API_KEY=your-key
uv run python run.py invoice_processing --limit 5

The default model is typesafe:jev-1.13.0; you can swap in OpenAI, Anthropic, Fireworks, Groq or Cerebras models with their own keys. Scores measure agreement with reference answers produced by large LLMs, not human labels. What those numbers show, and what they do not, is in our breakdown of TypeSafe AI's Jev benchmarks.

Open-Source Projects That Imitate Jev

Search GitHub for Jev and some of the most-starred results are not clients at all but attempts to reproduce the decision-model idea on open models you host yourself — projects such as SemIf (formerly OpenJev), NanoJev, AnyJev and simple-jev. By their own descriptions, they read option probabilities out of general open models or train small replicas. SemIf states that it is independent of TypeSafe, and none of them is published by TypeSafe.

They are a different trade, not a free Jev: the quality ceiling is whatever open model you point them at, and calibration is yours to measure. Our Jev vs OpenJev comparison covers when self-hosting beats a hosted API.

FAQ

Is Jev open source?

No. Jev is a proprietary hosted model. TypeSafe open-sources its SDKs, agent skill, adapter, eval code and n8n node under MIT or Apache-2.0, but not the model, its weights or its training code.

Is there an official Jev GitHub repository?

There is an official organization, typesafe-ai, but no repository for the model itself. The closest things to "the Jev repo" are the Python and JavaScript SDKs and the agent skill.

What is typesafe-ai/jev?

A URL people search for that does not exist — it returns a 404. If you land on a repo with a similar name under another account, it is not TypeSafe's.

Is github.com/jev-ai official?

No. It was created on September 20, 2026, has no link to TypeSafe, and its repositories are mostly one-line placeholders named after search terms, pointing to a third-party website.

Where are the Jev docs and cookbooks?

On docs.typesafe.ai, not GitHub. The docs include a cookbooks section with end-to-end recipes such as re-ranking, guardrails and function calling, and the agent skill can point your coding agent to them.

Can I self-host Jev from GitHub?

No. There are no weights to host. You can self-host open-model imitations, or cache and distill Jev's answers locally with community tools, but the model itself is only available as an API.

The Bottom Line

The real Jev GitHub is small and useful, and it is not where the model lives.

• TypeSafe's official organization is typesafe-ai: SDKs, an agent skill, an LLM comparison adapter, eval code and an n8n node, mostly MIT-licensed.
• There are no weights, no training code and no typesafe-ai/jev repo.
• TypeSafeAI calls itself unofficial; jev-ai is unaffiliated. Verify any repo against links on typesafe.ai before trusting it with a key.
• To try Jev, you need an API, not a clone.

The quickest path from reading to running: open the Jev AI playground, edit a scenario with your own text, and when the answers look right, send the same request from your code through the Jev AI API.

Sources

• TypeSafe on GitHub — The official organization and its repository list, checked through the GitHub API on September 30, 2026.
• Python SDK — TypeSafe AI docs — Install commands and the link from TypeSafe's docs to the SDK source.
• Agent skill — TypeSafe AI docs — Official install commands for the typesafe-ai/skills repository.
• Models — TypeSafe AI docs — Hosted-only model, shared weights and no per-customer fine-tuning.
• WorkflowEvals — TypeSafe AI on GitHub — Code and instructions for reproducing TypeSafe's published evals.

Repository counts, versions and star counts are a snapshot from September 30, 2026. Details about unofficial organizations come from their public GitHub profiles on that date. Community projects are not affiliated with TypeSafe unless TypeSafe links to them.