Vercel AI Gateway adds TypeSafe clients and HTTP API for Jev decision model

Vercel has added support for Jev to its AI Gateway, giving developers three ways to reach the model: through an existing TypeSafe client, through a plain HTTP endpoint, or through the AI SDK. According to the Vercel blog, requests made over any of these routes are billed via the gateway and appear next to other model traffic in its usage and observability views.

What Jev does

Vercel describes Jev, a model from TypeSafe AI, as a "probabilistic decision model for software." Rather than producing free-form text that a caller has to parse, it accepts a description of a state along with typed questions and returns answers that carry probabilities. Three question types are supported: boolean questions return a probability between 0 and 1, choice questions pick a single option from a set defined by the caller, and score questions rate something against a scale the caller defines.

The example Vercel publishes illustrates the intended use case: deciding whether an agent should take another step after it has fixed a checkout bug and its tests pass. That is the kind of control-loop judgement agent builders often handle today by prompting a text model and scraping the reply.

Three ways in

For teams already running a TypeSafe client, migration is deliberately small. Vercel says switching means pointing the client at the gateway base URL and supplying a gateway API key; existing systemOne calls, noul questions and response shapes remain exactly as they are.

The HTTP API opens the model to any language or framework. Callers send a POST request to the /v1/evaluate endpoint with the model identifier typesafe-ai/jev, a state string and a questions object, then read typed answers back from the response.

TypeScript developers can instead use the AI SDK, where an experimental evaluate function accepts the same model identifier, state and questions, and returns answers as typed objects with probability fields.

One bill, one view

Whichever route is used, billing runs through AI Gateway. Vercel frames this as the practical payoff: Jev calls sit alongside other model calls in usage reporting and observability, so evaluation traffic can be budgeted and inspected like any other inference rather than existing as invisible side traffic.

Jev in eve

Jev is also wired into eve, which Vercel describes as a framework for building and deploying agents with sandboxed compute, human approval steps and evaluation tooling built in. There, Jev acts as the default evaluation model, backing automatic model selection, typed evaluations and automated tool approvals.

Why it matters

The addition pushes AI Gateway beyond text-generating models into structured decision-making, a related but distinct job. Typed, probability-bearing answers remove the parsing step that makes text-model-based agent control loops brittle, and the three integration paths keep the barrier low: existing TypeSafe users need only a URL and key change, while the HTTP endpoint makes the model reachable from stacks Vercel does not natively package for. Unified metering and observability mean evaluation calls can finally be tracked and debugged like ordinary inference. For teams running agents in production, that combination — predictable response shapes, language-agnostic access and consolidated billing — targets the operational gaps that tend to surface once evaluation logic moves from prototype to production.