VercelEngineering PostOfficial Source

AI Gateway now supports TypeSafe clients and an HTTP API for Jev

Jev is now callable through TypeSafe clients, HTTP, or Vercel AI SDK, returning typed probabilistic decisions that agents can consume without parsing prose.

Vercel · Sep 21, 2026

Open Source Open Markdown Open JSON

Source Summary

Vercel AI Gateway now exposes Jev through **three integration paths**: an existing TypeSafe client, a language-agnostic HTTP API, and AI SDK. Jev returns typed boolean, choice, or score answers with probabilities rather than generated text.

Practical Implication

Use it where an agent needs a machine-readable decision boundary, such as continuation checks, model selection, evaluations, or tool approval. Existing TypeSafe users can migrate by changing the base URL and API key.

Agent-Ready Context

Vercel AI Gateway now exposes Jev through **three integration paths**: an existing TypeSafe client, a language-agnostic HTTP API, and AI SDK. Jev returns typed boolean, choice, or score answers with probabilities rather than generated text.

Use it where an agent needs a machine-readable decision boundary, such as continuation checks, model selection, evaluations, or tool approval. Existing TypeSafe users can migrate by changing the base URL and API key.

Jev is probabilistic, so a threshold such as **0.8** remains an application policy rather than certainty. The material supplies no calibration, quality, latency, or pricing results.

Connected Context · Feed7 Judgment

Jev adds a narrow, typed decision primitive alongside the gateway’s existing routing, transport, and evaluation infrastructure. It can simplify machine-enforced branches and acceptance checks, but its probabilities still require application thresholds and measured calibration; it does not replace access controls, trajectory-level evaluation, or deterministic production checks.

How Tailscale built a customer-facing model router on AI GatewayJev could supply structured routing decisions within this gateway pattern, while Tailscale’s identity checks and credential isolation remain separate enforcement controls.Run Terminal-Bench and other Harbor evals on Vercel SandboxHarbor’s isolated evaluation loop provides a way to test Jev thresholds and decision quality before using its probabilistic outputs in agent control flow.Mousepower: agents that can’t be measured, can’t be managed. — Maximillian Piras, YutoriJev’s booleans, choices, and scores can implement the cheap acceptance checks advocated by Mousepower, provided teams first define trustworthy outcome criteria.The Future of Evals: From LLM as a Judge to Agent as a Judge — Aparna Dhinakaran, Arize AIJev narrows evaluation to predefined typed questions, whereas agent-based judging addresses failures across variable trajectories; the approaches cover different evaluation scopes.

Context Map

benchmarkcoding#agent-evals#agent-reliability#gateways

Uncertainty

Jev is probabilistic, so a threshold such as **0.8** remains an application policy rather than certainty. The material supplies no calibration, quality, latency, or pricing results.