Azure connection
One Azure OpenAI resource, for AI steps to run on. In the connection picker it is Azure.
Powers these nodes
No Azure-specific nodes yet. It serves the built-in AI Agent action, the AI Judgment condition, and your default model—anywhere a Provider Account is chosen.
Fields
| Field | Required | Description |
|---|---|---|
| Name | Yes | A name for the account. It labels the account everywhere it is offered, so say whose account it is. |
| Endpoint | Yes | The resource endpoint from the Azure portal, for example https://my-resource.openai.azure.com—or one ending in the v1 path to reach the v1 API. |
| API Surface | No | Which request layout the resource is addressed on: Auto, Deployments, or V1. Leave it on Auto for a resource on one of Azure's own hostnames. |
| API Version | No | The Azure API version requests are addressed on. Leave it blank for the built-in default. Hidden on the v1 surface, where it means nothing. |
| Credentials | Yes | API Key—either of the two keys on the resource's Keys and Endpoint page. Stored encrypted and never logged. |
Name an API Surface only for a gateway that cannot be recognised from its hostname—a vanity domain or a private alias. Azure's own detection works from a fixed list of hostnames, so a resource that is not on one of them gets a request with nothing added to it, and every call comes back as a 404 with nothing to explain it. Naming the surface is the way out of that.
A newer API Version than the built-in default still works, so a resource can move forward without waiting on an upgrade here.
The deployment name is the model id
Azure routes on the name you gave your deployment, not on the model's own name—so unlike every other provider account, it publishes no model list to pick from.
Type the deployment name, found under Deployments inside the resource, into the Model field wherever you choose one. There is nothing to browse, and nothing has gone wrong when the list comes up empty.
Compatibility
Azure speaks the OpenAI protocol but not all of it, and the Compatibility tab carries two switches for that:
| Switch | Turn it on when | Default |
|---|---|---|
| Don't Use response_format | The endpoint cannot return an answer in a shape you asked for. Steps that need structured output ask the model to call a function instead, which works anywhere function calling does. | On |
| Don't Use response_format Together With Tools | The endpoint handles a shaped answer on its own but rejects a request that asks for one and offers tools. Only those requests switch approach. | On |
Both are on by default here, because whether they are needed depends on the model behind the deployment and the API version. Turn them off once you have confirmed the resource accepts a shaped answer—the symptom of leaving one wrongly on is nothing at all, and of turning one wrongly off is a step failing because the model returned no structured output, or the wrong shape.
Connecting
Credential-based: there is no Connect button. Enter the endpoint and key and save.