01
Overview
Jev's API is built around a different contract from a normal chat-completions API. Instead of asking a model to generate text, you send application state plus one or more typed questions. Jev returns structured decisions and probabilities that your code can evaluate directly.
This guide covers the current TypeSafe API and the Vercel AI Gateway paths available as of September 22, 2026.
02
Before you start
Jev is TypeSafe AI's System One decision model. Use it when the possible decisions are known in advance: route a request, choose an option, score a condition, decide whether to continue, or gate an action. It is not intended to write the user-facing response.
For open-ended text, keep a generative model in the workflow and use Jev at the constrained decision points.
03
Direct TypeSafe API
TypeSafe's public OpenAPI documentation exposes two core endpoints: GET /v1/models for model discovery and POST /v1/systemone for evaluations. Authentication uses an API key in the Authorization header as a Bearer token.
A System One request supplies shared state plus structured questions. The current API schema exposes Choice, Score and Noul-style question/answer types. TypeSafe's SDK terminology and Vercel's newer abstraction differ slightly, so follow the schema for the endpoint you are actually calling rather than assuming the names are interchangeable.
04
What should go in state?
State is the information Jev should evaluate. It can represent a support ticket, tool result, agent status, transaction summary, policy-relevant facts or other program context.
Keep state focused on the decision. A decision model is most useful when the surrounding application defines the workflow and the model only supplies the fuzzy judgment that would otherwise require brittle rules.
05
Define the decision before calling the model
The key design step is declaring the possible output space before inference.
Use a choice when Jev should select among named options such as billing, technical support or sales. Use a score when you need an ordered rating such as risk or urgency. Use a boolean/probability-style question when the application needs a threshold decision such as whether an agent should continue working.
This constraint is the reason Jev outputs can be consumed directly by software without parsing arbitrary prose.
06
Vercel AI Gateway: HTTP API
On September 21, Vercel added a native HTTP evaluation path for Jev. New integrations can POST to Vercel AI Gateway's /v1/evaluate endpoint and specify typesafe-ai/jev as the model.
The request includes the model, state and questions. Vercel documents boolean, choice and score question types. The response preserves structured answers and probabilities rather than returning a chat message.
This HTTP path is useful when the application is not written in TypeScript or does not use Vercel's AI SDK.
07
Vercel AI Gateway: AI SDK
Vercel AI SDK 7 added an experimental evaluate API for decision models. Jev can be selected as typesafe-ai/jev, then called with state and a question map.
The result exposes answers by question ID. TypeSafe-specific confidence metadata can also be available through provider metadata. Vercel recommends calibrating probabilities and confidence against labeled examples from the actual workflow.
08
Migrating a TypeSafe client to Vercel
Vercel also supports TypeSafe-compatible clients. An existing TypeSafe client can point at the Vercel AI Gateway TypeSafe base URL and use an AI Gateway key while retaining its existing systemOne calls and response shapes.
This means teams can choose between direct TypeSafe access and Vercel's gateway layer without rewriting the decision logic itself.
09
Example agent pattern
A practical agent loop can divide responsibilities this way:
1. A generative model interprets the request and proposes a next action.
2. Application code validates permissions and tool arguments.
3. Jev evaluates a constrained question such as whether the proposed action is safe, whether the task is complete, or which subagent should run next.
4. Code compares the returned probability with a threshold.
5. High-confidence cases continue automatically; uncertain cases can be escalated for review.
6. A generative model produces any final user-facing explanation.
Jev should not own permissions or irreversible side effects. Those remain application responsibilities.
10
Do not blindly use a 0.5 threshold
A probability is useful only if the threshold matches the cost of being wrong. A low-risk classification may tolerate more automation than a financial, security or destructive action.
Collect labeled examples from your own workflow, measure how Jev's probabilities behave, and choose thresholds for autonomous action, retry or human review accordingly.
11
Pricing
TypeSafe currently lists direct Jev input at $0.042 per million tokens and output as free. Requests routed through a gateway are billed according to that gateway's current terms and promotions, so verify the active provider price before estimating production cost.
12
Common mistakes
Do not use Jev when you actually need free-form generation. Do not treat schema-valid output as proof that the decision is correct. Do not give the model authority that belongs in code. Do not copy vendor benchmark speedups into capacity planning without testing your own state sizes and network path. And do not assume direct TypeSafe and Vercel abstractions use identical names for every question type.
13
FAQ
What endpoint does the direct Jev API use?
TypeSafe currently documents POST /v1/systemone for System One evaluations.
Can I use Jev without TypeScript?
Yes. TypeSafe has a direct HTTP API, and Vercel added a language-agnostic HTTP evaluation endpoint for Jev on September 21.
What Jev question types are available?
TypeSafe's API schema documents Choice, Score and Noul-style questions. Vercel's evaluation abstraction exposes boolean, choice and score.
Can Jev choose an agent tool?
Yes, tool or subagent selection is one of the documented decision-model use cases. Tool execution and permissions should remain controlled by application code.
Should Jev replace my LLM?
Usually not. Jev is designed for constrained decisions; a generative model remains useful for interpretation and user-facing generation.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.