Supported providers

A decision model entity connects the gateway to one provider. Every provider below answers the same System One requests: an application talks to the gateway, and the provider behind it is your choice.

Provider	provider	Default base URL	Example model
TypeSafe	typesafe	https://api.typesafe.ai/v1	jev-latest
OpenRouter	openrouter	https://openrouter.ai/api/v1	typesafe/jev-1.13
Cloudflare Workers AI	cloudflare	https://api.cloudflare.com/client/v4	@cf/cloudflare/clef-flash
Liquid AI	liquid	https://api.liquid.ai/decisions/v1	d1:free
Telnyx	telnyx	https://api.telnyx.com/v2/ai/typesafe/v1	telnyx/decision-flash
Prem AI	prem	https://gateway.prem.io/typesafe/v1	dgemma
Vercel AI Gateway	vercel	https://ai-gateway.vercel.sh/typesafe/v1	typesafe-ai/jev
Any System One server	systemone-compatible	yours	yours
One of your text providers	llm-emulation	—	the model of the text provider

Entity configuration

{
  "id": "decision-model_jev",
  "name": "Jev",
  "provider": "typesafe",
  "config": {
    "connection": {
      "base_url": "https://api.typesafe.ai/v1",
      "token": "${vault://local/typesafe-token}",
      "timeout": 30000
    },
    "options": {
      "model": "jev-latest"
    }
  }
}

Field	Description
config.connection.base_url	Where the provider is reached. The gateway calls <base_url>/systemone
config.connection.token	The API key of the provider. Several keys separated by commas are used in turn
config.connection.timeout	How long to wait for the answer, in milliseconds
config.options.model	The model used when a request names none
config.options.allow_config_override	true by default: a request can name another model. Set it to false to pin the model
fallback_ref	The decision model taking over when this one cannot answer. See fallback
fallback_model	The model asked to the fallback, its own default model when empty
models	Model constraints: include / exclude patterns of the models consumers may use, and require_known_costs to refuse a model the gateway cannot price

TypeSafe

The reference implementation of the System One API, with the Jev models. Pin a versioned model such as jev-1.13.0 when your thresholds must stay stable, or follow jev-latest.

OpenRouter

OpenRouter serves Jev and other decision models (Liquid d1, Inception Mercury Decide...) on its own System One endpoint. Two things come with it:

• the routing preferences of OpenRouter: a provider object in the request (order, allow_fallbacks...) is forwarded as it is
• the cost of each call, reported by OpenRouter itself and used as is for cost tracking and budgets, including for the models no price table knows

Cloudflare Workers AI

Clef and Clef-flash, the open-weight decision models of Cloudflare. They take the same requests as Jev, with up to 64 questions, and can also look at pictures through an images array.

{
  "provider": "cloudflare",
  "config": {
    "connection": {
      "account_id": "your-cloudflare-account-id",
      "token": "${vault://local/cloudflare-token}"
    },
    "options": {
      "model": "@cf/cloudflare/clef-flash"
    }
  }
}

The model is named the way Workers AI names it (@cf/cloudflare/clef, @cf/cloudflare/clef-flash).

Liquid AI, Telnyx, Prem AI, Vercel AI Gateway

These providers expose the System One API under their own base URL, with their own models. Create the entity with the matching provider and your API key: the base URL of the table above is used unless you set another one.

Self-hosted and compatible servers

systemone-compatible reaches any server speaking the System One API: an open-weight model served by Laya or vLLM, a LiteLLM proxy, an internal service.

{
  "provider": "systemone-compatible",
  "config": {
    "connection": {
      "base_url": "http://vllm.internal:8000/v1",
      "path": "/systemone",
      "provider_name": "Internal vLLM",
      "headers": {
        "Authorization": "Bearer {api_key}"
      },
      "token": "${vault://local/vllm-token}"
    },
    "options": {
      "model": "clef-flash"
    }
  }
}

path is appended to base_url (/systemone by default) and headers are sent with every call, {api_key} being replaced by the token.

To put a price on a self-hosted model, give the entity the costs-tracking-provider and costs-tracking-model metadata of the model it runs, or declare your own prices in the cost tracking settings.

Decisions from a text provider

llm-emulation lets one of your text providers answer System One requests. The gateway asks the chat model for the probability of every outcome and computes the answers — choice, score, confidence — exactly the way a decision model reports them. Your applications call the same endpoint and read the same response, whether a decision model or an LLM is behind it.

{
  "provider": "llm-emulation",
  "config": {
    "connection": {
      "provider": "provider_xxxxxxxxx"
    },
    "options": {
      "model": "gpt-4o-mini"
    }
  }
}

• connection.provider — the LLM provider entity answering the questions
• options.model — the model of that provider to use, its default model when omitted

The call is a call of the text provider: its guardrails, its model constraints and its budgets apply, and it is billed, counted and logged as such — once.

tip

The probabilities of an emulated decision are the ones the chat model states. They are a good way to adopt the decision API with the models you already have, and a reason to tune your thresholds again when you later move to a model trained to decide.

Pin the model

The TypeSafe SDKs always send a model name, jev-latest unless told otherwise. Set allow_config_override to false and the entity serves its own model whatever the request asks for:

"options": {
  "model": "@cf/cloudflare/clef-flash",
  "allow_config_override": false
}

Applications written for Jev keep running unchanged while the gateway serves Clef, a self-hosted model, or tomorrow's better one. The model becomes a deployment decision.

Fallback

A decision is made to be waited for. Give a decision model a fallback_ref and another one takes over when it cannot answer:

{
  "id": "decision-model_jev",
  "provider": "typesafe",
  "fallback_ref": "decision-model_clef",
  "fallback_model": "@cf/cloudflare/clef-flash"
}

• The fallback is used on technical failures: no answer (timeout, connection), rate limiting (429) and server errors (5xx, 529).
• A request the provider refused as invalid is not sent elsewhere: it would be refused there too.
• A low confidence is an answer, not a failure. Probabilities are not comparable between models, so the gateway never shops for a more confident one.
• The call is priced at the price of the model that served it. It counts against the budgets of both models, the one that was asked and its fallback, and once in a budget that has both in its scope.
• Fallbacks can be chained, and two models falling back on each other are each tried once.