Jev API and SDK guide: TypeSafe and OpenRouter

The direct TypeSafe request

The official quick start links to the console, API-key page and playground. The HTTP API accepts POST https://api.typesafe.ai/v1/systemone, with a bearer token and a JSON body containing state, questions and the selected model.

{
  "model": "jev-latest",
  "state": { "message": "The upload fails after I choose a PDF." },
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle message?",
      "criteria": {
        "technical": "Errors and broken product functionality",
        "billing": "Invoices, charges and subscriptions",
        "other": "None of those categories"
      }
    }
  }
}

This is a request-shape example, not a captured model result. Real calls require an authorized key and can incur charges. Keep credentials on the server or in your local environment rather than in browser bundles.

Official SDKs and community clients

The official Python package is typesafe-sdk; the JavaScript/TypeScript package is @typesafe-ai/sdk. Both documented examples read TYPESAFE_API_KEY. Installing a client library installs code that calls an API; it does not place the hosted model's weights on your CPU.

pip install typesafe-sdk
# Or, for a JavaScript/TypeScript project:
npm install @typesafe-ai/sdk

For Java, Rust, Swift, Go or a Unix pipeline, inspect the SDK directory and each owner's current documentation. SemDecide is a community CLI for semantic predicates, routing and scoring. A framework with type-safe text generation, such as TypeLLM, has a different role from a decision-only model. Package names and familiar branding do not establish official ownership.

OpenRouter, Google and Vertex: check the actual surface

OpenRouter's official classification cookbook uses https://openrouter.ai/api/alpha/decisions, OPENROUTER_API_KEY and typesafe/jev-1.13. It sends state and typed questions, then validates the returned answers. Check that documentation for current versions and provider-specific error handling.

The public OpenRouter model catalogue also describes typesafe/jev-router, which selects models for requests. That routing product is different from asking the Jev decision model a typed question. A normal chat-model selector therefore does not tell you which Decisions models or aliases are accepted.

Jev's developer is TypeSafe AI. A search for “Google Jev” or “Jev Vertex” does not establish that a native Jev model deployment exists on Vertex. Before connecting any provider, verify its documented model ID, endpoint, authentication and response shape. An adapter using a Google model to imitate the System One interface is still backed by that other model. For a hands-on walkthrough of the OpenRouter route, including the Decisions endpoint and the separate jev-router product, see Jev on OpenRouter.

What to verify before changing a base URL

• Does the client send the request shape the backend expects?
• Are Choice, Score, Noul, confidence and missing-field behavior compatible?
• What does a timeout, invalid answer, quota failure or rate limit do?
• Does the service accept an alias or require a versioned model ID?
• Which model actually answered, and can that identity be logged?

The official System One Adapter exposes a similar evaluation interface backed by LLM APIs. It is useful for controlled comparisons; interface similarity does not make the underlying model, calibration or cost identical.

A small first integration

Begin with one decision and a small labeled set of your own examples. Validate transport and response handling without granting the model authority over an external action. Then test ambiguous inputs, missing confidence, malformed responses and retries. Pin the version when comparing outcomes; the model documentation explains that aliases can move to new releases.