Microsoft Foundry configuration

Foundry is the default model provider. LlmResourceFactory.AddMicrosoftFoundryResource declares a Foundry resource and model deployment through Aspire. The agent service creates a local ChatClientAgent that calls that model.

Starting the local AppHost can create billable Azure resources. Confirm a development subscription, permitted region, model quota, provisioning permissions, and cleanup owner first. Chapter 0 covers the first-run checks. The learner starter remains cloud-free until model access is activated in the first-agent lesson.

Authenticate

Install the Azure CLI, then sign in and check the account:

Bash

# Bash
az login
az account show --query "{subscription:name, subscriptionId:id, tenant:tenantId}" --output table

PowerShell

# PowerShell
az login
az account show --query "{subscription:name, subscriptionId:id, tenant:tenantId}" --output table

Use az login --tenant YOUR_TENANT_ID and az account set --subscription YOUR_SUBSCRIPTION_ID when you need explicit selection. Replace the placeholders and keep account details out of published logs.

The completed Program.cs uses DefaultAzureCredential with the token scope https://cognitiveservices.azure.com/.default. In development it excludes the managed-identity probe so local developer credentials can be used. AZURE_TENANT_ID sets the credential tenant when present. This path uses identity-based authentication.

For a deployed application, check its managed identity and access to the actual model resource separately from local CLI authentication.

Model settings

The checked-in apphost.settings.json requests:

{
  "LlmProvider": "MicrosoftFoundry",
  "AgentMode": "HandOff",
  "MicrosoftFoundry": {
    "DeploymentName": "gpt-5-mini",
    "ModelVersion": "2025-08-07",
    "ModelFormat": "OpenAI",
    "SkuName": "GlobalStandard",
    "SkuCapacity": 100
  }
}

Confirm these sample values against model availability and quota in your subscription. Keep an explicit model version: omitting it makes the factory request version 1, which may be invalid for the selected model.

Aspire passes a chat connection string containing Endpoint and Deployment to the agent service. The Foundry branch uses those values to construct an OpenAI-compatible endpoint and register client.AsIChatClient().

During the workshop, WorkshopHosting.cs supplies this source-derived setup so the first-agent lesson can focus on the ChatClientAgent constructor and interview instructions. Learners activate the helper after writing the agent and retain it in the completed application. See workshop startup helpers.

Reuse an existing account and deployment

The standalone sample provisions by default. To reuse existing resources, set MicrosoftFoundry:UseExisting to true and provide all four values under MicrosoftFoundry:Existing: Name, ResourceGroup, SubscriptionId, and DeploymentName. The deployment name is the Azure resource name, often chat, rather than the model name in the normal provisioning settings.

For a local run, also set Azure:AllowResourceGroupCreation to false and configure the AppHost’s normal Azure:SubscriptionId, Azure:ResourceGroup, and Azure:Location context. Keep these account-specific values in local user secrets or deployment configuration. Model version, SKU, and capacity settings are ignored while reusing a deployment.

AddExistingMicrosoftFoundryResource submits an ARM deployment containing only existing account and deployment references. It reads the endpoint and verifies the deployment exists. It does not create an account, model deployment, capability host, or role assignment. This requires permission to evaluate the ARM deployment and existing permission to call the model. Invalid settings, missing resources, or denied access fail explicitly; there is no automatic provisioning fallback.

The workshop’s learner archives enable reuse from the start, with resource-group creation disabled. They remove the repository’s inherited UserSecretsId, so each file-based learner AppHost uses its own path-specific store. Use dotnet user-secrets set KEY VALUE --file ./apphost.cs from that learner folder; file-based apps do not support dotnet user-secrets init --file. Copy only the resource identifiers from Chapter 0. Do not copy its secrets file, cached Azure:Deployments values, or a ConnectionStrings:chat override.

Entry points and ownership

From the repository root, start the local file-based AppHost:

Bash

# Bash
aspire start --apphost ./apphost.cs

PowerShell

# PowerShell
aspire start --apphost ./apphost.cs

The complete application also needs the container engine for Cosmos and MarkItDown. Review any Azure context and provisioning prompts before proceeding.

The project-based AppHost is src/InterviewCoach.AppHost; it uses its own appsettings.json and is the deployment target in azure.yaml. In learner projects, prepare that separate AppHost only for optional deployment. The core course uses root apphost.cs throughout.

Record the resources created by each AppHost. Reuse means the example and learner depend on the same model account; keep it until both are finished. Cloud resources remain after aspire stop, and azd down covers its selected environment. Review each scope in the cleanup reference.