Microsoft Foundry

Configure Microsoft Foundry deployments and access them through TrueFoundry’s AI Gateway.

Microsoft Foundry provides one account for Azure OpenAI models and models from other publishers, including Anthropic, Cohere, Meta, and Mistral. The integration is generally available in TrueFoundry v0.175.0. It uses the OpenAI-compatible v1 surface for compatible deployments and provider-specific routes where required.

Moving an existing Azure OpenAI or Azure AI Foundry account is covered in Migrate to Microsoft Foundry. Hub and hub-based Azure resources cannot be migrated.

Before you start

You need:

• A Microsoft Foundry resource with one or more deployed models.
• The resource’s OpenAI endpoint.
• The project endpoint used for deployment discovery and provider-specific Foundry routes.
• An API key or one of the supported Azure identity authentication methods.
• A TrueFoundry API key for requests through the AI Gateway.

Add a Microsoft Foundry account

1. Open the Microsoft Foundry model account form

   In the TrueFoundry dashboard, open AI Gateway > Models and select Microsoft Foundry.

2. Configure the resource

   Enter an account name, the Azure OpenAI endpoint, and the Microsoft Foundry project endpoint.

   The OpenAI endpoint should be the resource URL, for example:

   https://my-resource.openai.azure.com
   

   The project endpoint should include the project path, for example:

   https://my-resource.services.ai.azure.com/api/projects/my-project
   

   

   Microsoft Foundry Endpoints

3. Configure authentication

   Configure authentication at the account level using an API key, certificate, client secret, or workload identity federation.

   See the guides for certificate authentication, client secret authentication, and workload identity federation.

4. Add or discover deployments

   Use model discovery when the project endpoint is available, or add deployments manually.

   A model integration uses:

   • Name: the integration name used in Gateway model IDs.
   • Model ID: the deployment name in Microsoft Foundry.
   • Foundation model: the underlying model used for catalogue metadata and routing.
   • Deployment type: standard, global, datazone_us, or datazone_eu.
   • Model types: the operations supported by the deployment.

Discover deployments

Model discovery reads deployments from the project endpoint. It uses the catalogue to infer model types and public cost metadata when the foundation model is known.

Discovery is stateless. Include the current integrations in the request so existing integration names remain stable and new names avoid collisions.

If the foundation model is not in the catalogue, the deployment is returned with inferred.model_types: true so you can select the model types manually.

Call a Microsoft Foundry model

Gateway model IDs use the format:

<provider-account-name>/<integration-name>

For example, if the account is my-foundry and the integration is gpt-4.1, use my-foundry/gpt-4.1 in the request body:

curl "$GATEWAY_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer $TFY_API_KEY" \
  -H "x-tfy-provider-name: my-foundry" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "my-foundry/gpt-4.1",
    "messages": [
      {"role": "user", "content": "Explain Microsoft Foundry in one sentence."}
    ]
  }'

The TrueFoundry API key authenticates the Gateway request. The Azure credential configured on the provider account authenticates the upstream Microsoft Foundry request.

Supported operations

Microsoft Foundry models can use the following Gateway operations when the deployment and model integration declare the corresponding model type:

• Chat completions and legacy text completions
• Responses and Responses compaction
• Embeddings
• Image generation and image editing
• Text-to-speech, audio transcription, and audio translation
• Realtime WebSocket sessions
• Files, batches, and fine-tuning
• Cohere rerank
• Mistral Document AI OCR
• Anthropic Messages and Messages token counting for Claude deployments

OpenAI-compatible deployments

OpenAI-compatible deployments use the /openai/v1 upstream surface by default. The deployment name is sent as the upstream model value.

If a deployment does not support v1, set api_version on that model integration. TrueFoundry then uses the deployment-scoped Azure route with that API version.

Claude deployments

Claude deployments use the native Anthropic Messages route. The provider account must include the project endpoint so TrueFoundry can reach the resource-level Anthropic surface.

Use a Gateway model ID such as:

my-foundry/claude-opus-4-8

Cohere rerank deployments

Cohere rerank deployments use the Foundry Cohere route. Select rerank as the model type and use the deployment name as the integration’s model_id.

Mistral Document AI OCR

Select ocr as the model type and call the existing Gateway OCR proxy. TrueFoundry derives the upstream Mistral OCR route from the provider account’s project_endpoint.

import base64
import requests

with open("document.pdf", "rb") as document:
    encoded_document = base64.b64encode(document.read()).decode("utf-8")

response = requests.post(
    "{GATEWAY_BASE_URL}/proxy/ocr",
    headers={
        "Authorization": "Bearer your-truefoundry-api-key",
        "Content-Type": "application/json",
    },
    json={
        "model": "my-foundry/mistral-document-ai",
        "document": {
            "type": "document_url",
            "document_url": f"data:application/pdf;base64,{encoded_document}",
        },
        "include_image_base64": True,
    },
)
response.raise_for_status()
print(response.json())

Realtime models

Realtime models use the Gateway WebSocket endpoint:

ws://<gateway-host>/live/<provider-account-name>?model=<integration-name>

For example:

wss://gateway.example.com/live/my-foundry?model=gpt-realtime

The provider account is already present in the URL, so the realtime query parameter is the integration name rather than the full Gateway model FQN.