Jev 1.13 (TypeSafe): What a System One Model Can Do With MCP

What Jev 1.13 is

Jev 1.13 is a decisions model released on 18 September 2026. You send it a state object and a set of typed questions. It returns an answer per question: a choice with a probability for each option, a yes/no probability, or a score. It never produces prose or reasoning. OpenRouter serves it on a separate endpoint, POST /api/alpha/decisions, and rejects it on /chat/completions. The state is capped at 32K tokens.

What TypeSafe says

TypeSafe positions Jev for routing, classification, verification and ranking: decision points where a typed answer matters more than prose.
OpenRouter’s docs list ticket triage, agent oversight, structured extraction, gating risky tool calls and verification cascades as use cases.

These are the vendor’s own claims, not measurements of ours.

How it reaches MCP

Not as a tool caller. Jev has no function-calling interface and cannot be selected in Agent Studio.
As an eval judge. In MCP Playground evals, Jev can grade each agent answer against the expected one in place of an LLM judge.
As a gate you build yourself: ask Jev whether a pending tool call is safe before your agent executes it. This is one of OpenRouter’s documented use cases, and it happens outside the MCP protocol.

What to watch for

It returns numbers, never reasons. When Jev fails an eval answer here, the explanation says what probability it assigned and nothing more. Re-grade with an LLM judge if you need the verdict argued.
One state per question works better than one large state. Jev’s documented weak spots are indirection and large, noisy states, so the grader here makes one call per answer instead of batching them.
The 32K state limit is small next to chat models. Long tool transcripts have to be trimmed before they fit. The grader here sends at most 4,000 characters of each answer.

Run an eval against a mock server, then grade the same run twice: once with an LLM judge and once with Jev. Where the two disagree, the answer is borderline, and those are the cases your expected answers most need to pin down. All the mock servers →

Available here

Model	Model ID
Jev 1.13	`typesafe/jev-1.13`

Go deeper

Frequently asked questions

Can Jev 1.13 call MCP tools?

No. It is a decisions model: it answers typed questions with probabilities and has no tool-calling or text-generation interface. OpenRouter rejects it on the chat completions endpoint. It is not offered in Agent Studio for that reason.

How does MCP Playground use Jev?

As an eval judge. For each item it sends the task, the expected answer and the agent’s answer as the state, and asks two yes/no questions: does the answer convey what was expected, and does it admit failure. The first decides the verdict, with 0.5 as the cut-off. The second only sharpens the explanation.

Why use Jev instead of an LLM judge?

Calibration and speed. Each verdict is a probability rather than a yes/no, so you can see how confident the grade was, and there is no generated text to wait for. The trade-off is that it gives no reasoning.

What does System One mean?

TypeSafe’s name for models that make fast, structured decisions for software. They return a typed choice instead of free-form text, which borrows the fast-versus-deliberate framing of System 1 and System 2 thinking.

Jev 1.13