Skip to main content
TypeSafe provides System One models such as Jev that evaluate a piece of content against typed questions and return calibrated answers. TrueFoundry connects to TypeSafe natively through the AI Gateway. TypeSafe models use the decision model type. Call them on POST /v1/systemone (System One body) or POST /v1/decisions (OpenAI Decisions body). The AI Gateway translates between the two, so either route works. Full request and response schemas: Decisions API.
You can also reach TypeSafe through Custom Endpoints. The native integration below provides model access control and cost tracking, with models addressed as account-name/model-name. Prefer this page unless you specifically need a transparent proxy.
Jev is also available through OpenRouter, so you can call it with an OpenRouter API key instead of a TypeSafe one. Auto Routing can use a Jev model from either account as its classifier — see Auto Routing.

Adding Models

This section explains the steps to add TypeSafe models and configure the required access controls.
1

Select TypeSafe as the provider

From the TrueFoundry dashboard, go to AI Gateway > Models, click Add Model and choose To New Provider Account. Select TypeSafe from the provider list.
Add Provider Account page listing the supported model providers

Choose a provider in Add Provider Account

2

Add TypeSafe Account Details

Give a unique name to your TypeSafe account and complete the form:
  • API Key: your TypeSafe API key. It is sent to TypeSafe as an Authorization: Bearer token.
  • Base URL (optional): defaults to https://api.typesafe.ai. Point this at a Cloudflare account (https://api.cloudflare.com/client/v4/accounts/<account_id>) to serve System One models through Cloudflare Workers AI instead.
Under advanced settings, you can optionally set a Discount % that is applied to the upstream list price for this account.
TypeSafe Configure Account form in the provider account setup drawer

Configure Account for TypeSafe

3

Add Models

Click Add Model Manually and fill the form:
  • Display Name: the name shown in the TrueFoundry UI, for example jev.
  • Model ID: the TypeSafe model identifier, for example jev-1.13.0 or jev-latest. Pin a versioned ID so confidence thresholds stay stable across releases. On a Cloudflare Workers AI base URL, use the Workers AI model name instead (typesafe/jev, @cf/cloudflare/clef, or @cf/cloudflare/clef-flash).
  • Model Types: Decision (default).
  • Cost (optional): TypeSafe bills input tokens only, so set the output cost to 0.
4

Save and grant access

Click Test Credentials to check the account against the selected models, or click Save without Testing. To keep the account in Git instead, click Apply using YAML and apply the generated spec with tfy apply.To let other users, teams or virtual accounts use or manage these models, open Manage Accounts, click ⋮ on the account and choose Access Control > Grant Access. See Access control.

Inference

Call the model with the TrueFoundry model ID in the format your-typesafe-account/your-model-display-name. Send the content in state and named questions in questions. This example asks a yes/no (noul) question about a support message. Set GATEWAY_BASE_URL to your gateway base URL without a trailing slash and TFY_API_KEY to your TrueFoundry API key.
The response contains the versioned model that served the request, an answers object keyed by your question IDs, and usage with input_tokens and output_tokens. The same questions on POST /v1/decisions return OpenAI-shaped answers. See Decisions API for both routes, question types, and translation limits.

Cloudflare Workers AI

TypeSafe Jev and Cloudflare Clef also run on Cloudflare Workers AI. Point the TypeSafe account Base URL at your Cloudflare account and use a Workers AI Model ID: The AI Gateway unwraps Cloudflare’s {success, result} envelope so the response has the same answers / usage shape as api.typesafe.ai. Clef accepts an extra images field on the System One body; Jev does not. You still call {GATEWAY_BASE_URL}/v1/systemone (or /v1/decisions) with the TrueFoundry FQN. Do not call /ai/run yourself.