The hosted model is not. TypeSafe publishes the API, the docs, and the SDKs. Other people have trained or adapted their own decision models and published those.

Can I run Jev locally?

Not Jev's weights. You can run a different model that speaks the same request shape. Kev, Laya, and Von document a local POST /v1/systemone server. Pointing a client there does not make the answers Jev's.

What is the official Jev GitHub?

There is no official model repository on the pages we verified. typesafe-ai/skills is an official agent skill for the hosted API. Community repos such as browser-use/jev-ultrafast call that API. Decision-model repos are independent.

Jev GitHub and Open Source

TypeSafe has not published Jev weights. Official clients are the SDKs. Separate open models now serve the same /v1/systemone shape.

Searches for Jev GitHub, Jev open source, and Jev local want a repo to clone. Two different things now show up in those results.

What is official

Artifact	Where	Role
HTTP API	`https://api.typesafe.ai/v1/systemone`	The hosted model
Python SDK	PyPI `typesafe-sdk`	Official client
JS / TS SDK	npm `@typesafe-ai/sdk`	Official client
Agent skill	typesafe-ai/skills	Teaches a coding agent to call the hosted API
Docs	docs.typesafe.ai	Contract

SDK source is client source. It is not the Jev model.

This site is also on GitHub (picsai288/JevAIHub). It is an independent content repo, not TypeSafe.

What other people have published

The request shape is public, so a local server can accept state, model, and questions and return answers. On 2026-09-30 these READMEs documented that path:

Project	Repository	What the README says
Kev	jaredpalmer/kev	TypeSafe's Python SDK works against the local server unchanged
Laya	NandhaKishorM/laya	`laya.serve` exposes the router on `POST /v1/systemone`
Von	wfzyx/von	`von serve` exposes `/v1/systemone`

A second group never claims to be a base-URL swap. SemIf and AnyJev read probabilities out of an open model you already have. open-alternative-jev says in its README that it is a Python package, not a drop-in HTTP API. NanoJev publishes a 0.6B checkpoint and a game-task training loop.

Profiles, star counts, and the claim split (what the author measured versus what this hub re-ran) are on Open-source System One models and under Projects.

Tools that still call hosted Jev

These are local processes. The model stays on TypeSafe's API:

browser-use/jev-ultrafast, a browser agent. Overview: Jev Ultrafast.
doronp/jevc, which compiles policy prose into questions plus a code reducer.
jomatsu/pi-jev-auto-mode, a fail-closed tool-call gate.
ishantanu/jevmetrics, an alpha OpenTelemetry Collector processor.

When a local server is the right search

You need the data to stay on a machine you control, you need to fine-tune a decision model on your own labels, or the network round trip is too slow for the loop. Hosted Jev is still the model TypeSafe calibrates and versions. A compatible server is a different set of weights. Thresholds you tuned on jev-1.13.0 do not move with the client.

When GitHub will not help

You want Jev's weights, or a fine-tune of the hosted model. Official models docs say Jev is not customer-fine-tuned. We still have not found a weight drop from TypeSafe.

Common mistakes

Installing a similarly named package that is not typesafe-sdk / @typesafe-ai/sdk when you meant the hosted API.
Treating Kev, Laya, or Von as "Jev, self-hosted."
Treating a community MCP or agent repo as the official model repository.
Copying an author's accuracy table into a launch decision. Those tables are on the author's data.

Next: Open-source System One models or How to use Jev.