Is Jev open source?
The short answer: the Jev model itself is a hosted, closed-weight System One model from TypeSafe AI, so there are no weights to download. The tooling around it — an MCP server and client integrations on GitHub — is open, and you reach the model over a simple API.
People searching for Jev open source usually want one of two things: the model weights to run locally, or the GitHub repos that let an agent call Jev. The weights are not published — Jev is served from TypeSafe AI's endpoint, the same way most calibrated models ship. What is open is the layer you build with, so you never have to reverse-engineer the protocol.
What is on GitHub
The Jev MCP server (github.com/codaaiteam/jev-mcp) is open and lets any agent — Claude, Cursor, or your own loop — call Jev's classify, score, check, gate and decide primitives as tools. You can read the source, fork it, and run it yourself against your own key. Client snippets for the hosted endpoint are open too, so wiring Jev into an app is copy-paste, not guesswork.
- jev-mcp: an open MCP server exposing Jev to agents (install with npx github:codaaiteam/jev-mcp — set your jv_live_ key and it routes to the hosted gateway automatically, no extra config)
- Open client examples for the POST /api/v1/decide endpoint
- Closed-weight model: the System One weights are hosted, not downloadable
- Instant access — grab a jv_live_ key and call the API today
Open-source Jev-like projects
The model weights are closed, but a small ecosystem of open, Jev-like decision projects has appeared — people rebuilding the typed-decision idea on open weights so it can run locally. They vary a lot in maturity, hardware target, API compatibility and whether they accept image input, so treat this as a snapshot, not an endorsement. The most-searched right now is Laya (including Laya-MLX, which runs in low memory on Apple Silicon).
| Project | Runs | Weights | API-compatible | Notes |
|---|---|---|---|---|
| TypeSafe Jev | Hosted API | Closed, pinned | — | The reference: calibrated, versioned System One |
| Laya / Laya-MLX | Local (Apple Silicon, GPU) | Open | Partial | Most-searched open route; low-memory MLX build |
| OpenJev builds | Local GPU | Open | Some claim wire-compatible | Multiple independent implementations; a few add image input |
If you want the full cloud-versus-local comparison — hardware, calibration, cost and setup side by side — see Jev vs Laya. If you specifically want to keep decisions on your own machine, running Jev locally covers what is and isn't possible today.
Why the model is hosted, not open-weight
Jev is trained with RLCD to return calibrated probabilities, and calibration is only meaningful against a pinned, versioned model. Hosting keeps that version stable so the confidence numbers your thresholds depend on do not drift under you. It also means you get roughly 70–500ms latency and pay about $0.001 per decision — billed by input tokens at $0.42 per million, output free — without standing up any inference of your own.
If you need local inference
There is no open-weight Jev build to self-host today, so if you are looking at running Jev local you are really choosing between the hosted API and a heavier open LLM you would have to calibrate yourself. For typed, low-latency decisions the hosted endpoint is the practical path: POST https://jevtypesafeai.com/api/v1/decide with a jv_live_ key, or the official POST https://api.typesafe.ai/v1/systemone.
See also: Jev AI overview · Jev vs Laya · Running Jev locally · Jev ecosystem & tools · Playground
Related: Jev vs Laya · Jev local · Jev models
Get a key instantly
The weights are closed, but access is open — grab a jv_live_ key and call the API now.