Azure AI Foundry

Connect Azure-hosted OpenAI-compatible endpoints to Fabric Agents using Microsoft Entra ID (Azure AD) Bearer-token auth. Resource discovery, deployment selection, and token refresh are handled automatically.

Azure AI Foundry exposes Azure-hosted OpenAI-compatible endpoints. Fabric Agents talks to Foundry via the Pi SDK and authenticates with Microsoft Entra ID (Azure AD) Bearer tokens — not the older api-key header — so deployments behind tenant-scoped IAM work without manual key rotation.

When to use this provider

Your organisation requires Azure AD authentication for AI workloads.
You have a Foundry resource with one or more model deployments under it (e.g., gpt-5.2-codex, gpt-5.1-codex-mini, deployed Llamas, custom finetunes).
You want token refresh + deployment discovery handled by the app instead of curating a static api-key.

If you only have an api-key for the older Azure OpenAI Service (not Foundry), pick Azure OpenAI in the provider list instead — same Pi-SDK plumbing, simpler auth.

Prerequisites

A registered Azure AD app with User.Read and a Cognitive Services scope (https://cognitiveservices.azure.com/.default).
The app's tenant ID and client ID.
Permission on the target Foundry resource (Reader or higher).

If your org uses device-code OAuth, no client secret is needed; the desktop app handles the flow inline.

Connect in Fabric Agents

Open Settings → AI → Connections → Add.
Choose Azure AI Foundry from the provider picker.
Paste your Tenant ID and Client ID. (App registration must be configured for public-client / device-code flow.)
Click Sign in with Microsoft — Fabric Agents opens the device-code dialog. Complete the OAuth dance in your browser.
After token exchange, Fabric Agents lists your Foundry resources. Pick the one you want.
The app discovers deployments under the resource and populates the model picker. Save.

Subsequent launches refresh the access token in the background — no re-sign-in unless your refresh token is revoked or the app is uninstalled.

What gets stored

Per the credentials flow, the connection holds:

tenantId, clientId, resourceId, resourceName
accessToken, refreshToken, expiresAt (encrypted via the OS keychain where available)
The discovered model deployment list

Tokens are refreshed on the silent path before each request — if a 401 surfaces despite a fresh token, Fabric Agents marks the connection as needing reauth and prompts inline.

Selecting a deployment

Foundry deployments are named instances of a base model. The picker shows the deployment id (what you set when you created it in the Foundry portal) and the underlying base model. A session locks to the connection on first message but can switch deployments freely within the connection.

Connection failures

Error	Likely cause
`AADSTS65001: ... has not consented to ...`	The user/admin hasn't consented to the Cognitive Services scope. Ask your AD admin to grant tenant-wide consent for the app.
`403 Forbidden` discovering resources	The signed-in user doesn't have Reader on any Foundry resource in the tenant.
`404 deployment not found` on send	Deployment was deleted in Azure Portal but cached locally — reopen the connection, click Refresh deployments.
`invalid_grant` on token refresh	Refresh token expired or revoked. Click Reconnect to re-run the OAuth flow.

Providers overview — the full provider matrix
LLM connection schema — azure_ai_foundry field reference
Custom OpenAI-compatible endpoints — alternative if you front Foundry with a gateway