Jev in TanStack AI: ask typed questions

Yes. TanStack AI has an official adapter for Jev, a package called @tanstack/ai-typesafe. It lives in the TanStack/ai repository, uses the MIT licence and was first published on 18 September 2026, three days after Jev launched. Its version 1.0.0 came out on 22 October 2026, together with TanStack AI 1.0.0.

The adapter does one job. TanStack's page says it calls Jev through decide(), which returns a typed answer about your data, and that the adapter has no chat. That fits what Jev is: a decision model that returns choices, scores and probabilities, and writes no text.

• What is Jev?
• How to use Jev

Add the core package and the adapter. TanStack's README uses pnpm; its docs show the same two package names for npm, yarn and others.

pnpm add @tanstack/ai @tanstack/ai-typesafe

Then put your TypeSafe API key in the environment of your server. The adapter reads it when it runs.

export TYPESAFE_API_KEY="your-typesafe-api-key"

If you would rather pass the key in code, use createTypesafeDecider with the model and the key. TanStack's adapter page also lists three optional settings: a base URL (the default is https://api.typesafe.ai), extra headers, and your own fetch function. They let you send requests through a gateway or a proxy.

Call decide() with an adapter, the data to judge (called the state) and a map of questions. This is TanStack's own example, a support ticket routed three ways:

import { boolean, choice, decide, score } from '@tanstack/ai'
import { typesafeDecider } from '@tanstack/ai-typesafe'

const result = await decide({
  adapter: typesafeDecider('jev-latest'),
  state: ticket,
  questions: {
    queue: choice({
      instructions: 'Which team should handle this ticket?',
      options: {
        billing: 'Payments, invoices, refunds',
        tech: 'Bugs, outages, integrations',
        sales: 'Pricing, upgrades, new accounts',
      },
    }),
    urgency: score({
      instructions: 'How urgent is this ticket?',
      levels: ['low', 'medium', 'high'],
    }),
    refund: boolean({
      instructions: 'Is the customer asking for a refund?',
    }),
  },
})

Each key under questions becomes a field on the result, so result.queue.value is typed as 'billing', 'tech' or 'sales'. Your code can then branch on it in plain TypeScript. Jev still answers with a probability, and your code decides what to do with it.

• How to write Jev questions that give useful answers

TanStack's Evaluate guide lists the fields for each question type:

• choice. value is the option Jev picked, probability is the chance of that option, probabilities is the full map, and confidence is also present.
• score. value is the nearest level label, score is the raw fraction between the levels, and legend and probabilities are included. You must pass at least two levels, lowest first.
• boolean. value is true when the probability of yes is 0.5 or more. It has a probability and no confidence field. On TypeSafe's own wire this question type is called noul.
• meta. result.meta holds the model that answered and the token usage. The name meta is reserved, so you cannot use it as a question key.
• state. It can be a string, an object or an array. TanStack says an array counts as one state, not as a batch of separate items.

• Jev confidence routing
• Jev intent routing

On your server, never in the browser. TanStack's guide says decide() needs your API key, so you put the call in a route and have the page call that route. The tutorial builds this as a TanStack Start route in about ten minutes, with a client that posts the ticket and branches on the answers.

TanStack's own React example app, examples/react/evaluate in the repository, has a dropdown that runs the same questions through five adapters, which is a quick way to compare providers on your own tickets.

• Jev vs Laya: pay per call or run it yourself?

Since the 1.0.0 release, the error thrown for a failed TypeSafe request carries the HTTP status code in a status field. That lets you tell a rate limit (429), which deserves a retry, from a bad key (401), which does not, without reading the message. This is TanStack's own example:

try {
  await decide({ /* adapter, state, questions */ })
} catch (error) {
  if (error instanceof Error && 'status' in error && error.status === 429) {
    // Rate limited. Wait, then call decide() again.
  }
}

To cancel a call, pass an abortSignal. To watch usage, pass middleware: TanStack describes it as observe-only hooks for usage, finish, error and abort, and it supports OpenTelemetry spans.

decide() also throws before sending anything if the questions map is empty or a question is named meta. And if Jev answers but an answer cannot be mapped, for example because it is missing or has no probability for its value, it throws after reporting the token usage of that call.

Yes. TanStack says every evaluate adapter uses the same decide() call, so changing provider means changing the adapter. Its Evaluate guide lists these:

• TypeSafe directly: typesafeDecider('jev-latest') with TYPESAFE_API_KEY.
• OpenRouter: openRouterDecider('~typesafe/jev-latest') with OPENROUTER_API_KEY.
• Vercel AI Gateway: vercelGatewayDecider('typesafe-ai/jev') with AI_GATEWAY_API_KEY.
• Cloudflare: cloudflareDecider('typesafe/jev') with a Worker binding, or an account ID and API token.
• OpenAI: openaiDecider('gpt-6-luna') with OPENAI_API_KEY. TanStack notes that OpenAI can refuse a question, which makes decide() throw a refusal error.
• Ollaya, run locally: ollayaDecider('laya:latest'), which needs no key and talks to a server on your own machine.

• Jev on OpenRouter: model IDs and the Decisions API
• Jev vs OpenAI Decisions API

Things the documentation leaves open or says plainly:

• Model IDs. TanStack's page names jev-latest and jev-1.13.0 and says other TypeSafe model IDs can be passed as a string. TypeSafe's models page says jev-latest now points to jev-1.14.0. If you tuned thresholds against one version, send that version's ID instead of the alias, as TypeSafe advises.
• What it covers. TanStack documents three question types here: choice, score and boolean. Its page does not mention Box questions or images.
• A young library. TanStack AI reached 1.0.0 on 22 October 2026, and its 1.0.0 notes describe many removed APIs. Pin a version and read the upgrade guide before you move to a new major release.
• Cost. TanStack's pages give no price. You pay TypeSafe, or the provider you choose, at their own rates.

• Jev pricing
• Jev cost calculator