Feature Request: Add Native Support for Azure AI Foundry (Anthropic)

[ameyer117]

Is your feature request related to a problem? Please describe.
Microsoft and Anthropic have very recently announced a strategic partnership bringing Claude models to Azure AI Foundry (formerly Azure AI Studio).

While Pydantic AI currently offers first-class support for the standard Anthropic API, as well as cloud-specific integrations for AWS Bedrock and Google Vertex AI, there is currently no native support for Azure-hosted Anthropic models. Users wishing to use Claude on Azure must currently build custom providers to handle Azure-specific endpoints and authentication (Entra ID), which creates unnecessary friction for enterprise users within the Microsoft ecosystem.

Describe the solution you'd like
I would like to see a new model integration added for Azure, following the same pattern currently used for Bedrock and Vertex AI.

Specifically:

1. Add an implementation (e.g., AnthropicAzureModel) that wraps the AnthropicFoundry client recently added to the official anthropic Python SDK.
2. Ensure the implementation supports Azure-specific authentication methods, particularly Azure AD / Entra ID (via azure-identity), similar to how other Azure integrations are handled.
3. Maintain API consistency with the existing AnthropicBedrockModel and AnthropicVertexModel classes to ensure a uniform experience across different cloud providers.

Describe alternatives you've considered
The current alternative is manually instantiating the AnthropicFoundry client and wrapping it in a custom generic model adapter, which is verbose and prone to configuration errors compared to a native Pydantic AI implementation.

Additional context

• Announcement: Microsoft and Anthropic announce strategic partnership
• SDK Support: The official anthropic Python library has added the AnthropicFoundry client to support this integration directly. More info here. Pydantic AI should leverage this upstream capability.

[DouweM]

@ameyer117 AnthropicProvider already has an anthropic_client argument, that is currently type hinted to take AsyncAnthropicBedrock and AsyncAnthropicVertex clients, but should work with AsyncAnthropicFoundry as well. Have you tried if that works? If it it does, a contribution to update the dependency and add the type to the union would be appreciated!

[mathieu-junia]

I just stumbled on this issue, I've tried it and it works well. Whether it is with api_key or Azure Entra ID (provided you do have the right permission, eg Cognitive Services User).

This code worked for me (pydantic-ai==1.20.0)

from anthropic import AsyncAnthropicFoundry
from pydantic_ai import Agent
from pydantic_ai.models.anthropic import AnthropicModel
from pydantic_ai.providers.anthropic import AnthropicProvider

# -------
# API KEY
# -------

# Create client with API key authentication
client = AsyncAnthropicFoundry(
    base_url="https://your_ai_foundry_resource.services.ai.azure.com/anthropic",
    api_key="your_api_key",
)

model = AnthropicModel(
    "claude-sonnet-4-5", provider=AnthropicProvider(anthropic_client=client)
)
agent = Agent(model)
response = agent.run_sync("good morning")
print(response.output)

# ---------
# AZURE AAD
# ---------
from azure.identity.aio import (
    get_bearer_token_provider,
    DefaultAzureCredential,
)

token_provider = get_bearer_token_provider(
    DefaultAzureCredential(), "https://cognitiveservices.azure.com/.default"
)

# Create client with Entra ID authentication
client = AsyncAnthropicFoundry(
    azure_ad_token_provider=token_provider,
    base_url="https://your_ai_foundry_resource.services.ai.azure.com/anthropic",
)

model = AnthropicModel(
    "claude-sonnet-4-5", provider=AnthropicProvider(anthropic_client=client)
)
agent = Agent(model)
response = agent.run_sync("good morning")
print(response.output)

[DouweM]

@mathieu-junia Thanks for verifying! Would you be up for submitting a PR to bump the Anthropic SDK and add that type to the type hint? It would also be good to have some docs on support for those 3 platforms.

[mathieu-junia]

@DouweM sure, I can do that.