Jev — a decision model your agents can call

What it is

Jev is TypeSafe's decision model. It generates no text at all: you hand it a situation plus one or more typed questions, and it returns each answer with the probability it assigns. That makes it the right tool for the work that never needed prose in the first place — sorting items into buckets, picking a route, extracting a value, rating something against a rubric, or judging whether another model's output is good enough to act on.

Because the answer arrives with a confidence, you can gate on it: act above a threshold, escalate below it.

Where to find it

Jev has no screen and nothing to switch on. It reaches your agents as a skill an AI can use plus a small command-line tool the agent runs — there is no server and nothing left running between calls.

You see its effect in two places: the answers your agent gives you when it used Jev to decide, and the spend it adds to your gateway credit, which you can read in Settings → Plan & Usage.

How it behaves

How an agent uses it

Jev ships as a skill plus a small command-line tool — no server, nothing left running. An agent that has the jev skill runs one command per question and gets JSON back:

node scripts/jev-cli.mjs check "the deploy is safe to roll forward"
node scripts/jev-cli.mjs choose --state-file /tmp/situation.txt \
  --instructions "Which bucket?" --option bug="broken now" --option feature="new capability"
node scripts/jev-cli.mjs decide --request-file /tmp/questions.json      # several in one call

Nothing needs to be turned on and no key needs to be set: like Omniscio's other outside-API calls, Jev goes through Omniscio's gateway on your account, and each call is metered to your prepaid credit — the same credit that powers the built-in voices and the other gateway features. You can see the spend in Settings → Plan & Usage.

The gateway reaches Jev by two routes, and it tries the first one first: Omniscio's own TypeSafe account, and then the OpenRouter route as a backup. A call stops at the first route that answers, so backup capacity is only spent when it is actually needed. You do not choose between them and there is nothing to configure — from your side it is one call and one meter.

If you have your own OpenRouter key saved (Settings → My API Keys), it is used only when both gateway routes are unavailable — never as the first choice, and never while a gateway route could still answer.

What it costs

Jev is metered to your gateway AI credit at TypeSafe's rate, and every call reports its own exact cost. Every call — from an agent, the spam filter or the email safety screen — is also added to your AI spend in Stats → Spend and shows as its own Jev line in the daily AI spend report, and the first time Jev spends more than $1 in a day you get one inbox card saying so. Typical calls land around a hundredth of a cent. Ask several questions in one call rather than several calls — extra questions in the same call cost almost nothing extra.

Reading the answer

Every answer comes with a confidence between 0 and 1.

• High confidence — act on it.
• Low confidence — the model is telling you it is not sure. Escalate, ask a narrower question, or hand the decision to a person.

A confidence is a judgement, not a verified fact. Never let an answer — however confident — be the only thing standing behind an irreversible action like deleting something, sending a message, or merging code.

When Jev says it cannot answer

If no one can be asked — you are signed out, your gateway credit is used up, or the shared route to a provider is unavailable — Jev says so plainly and tells you which of those it is. That is deliberately not the same as an answer of "no": an agent that cannot ask you will report that it could not ask.

What Jev is not

• Not a chatbot. It writes no prose, explains nothing, and summarises nothing.
• Not a browser agent. This is the decision model, not TypeSafe's browsing tool.
• Not for anything with steps. One decision per question, phrased so it can be answered.

Related

Other pages about an AI being given a packaged job to do: the Use skills page for how an agent picks one up, and Cost control for the gateway credit Jev's calls are metered against.