Azure AI Foundry

AzureAIFoundry is Agno's classic adapter for Azure-hosted models. Its azure-ai-inference dependency retired on August 26, 2026. Existing endpoint availability is a separate question; use the current API for new projects.

Current Foundry API

For a deployment supporting Chat Completions on the current Foundry API, use OpenAILike. Copy the resource's OpenAI-compatible base URL (ending in /openai/v1), its key, and the exact deployment name from your Foundry project. Check the deployment's supported operations before choosing a model. See Microsoft's current inference example.

uv pip install -U agno openai
export FOUNDRY_API_KEY="your_resource_key"
export FOUNDRY_BASE_URL="https://your-resource.services.ai.azure.com/openai/v1"
export FOUNDRY_DEPLOYMENT="your_chat_deployment"

foundry_agent.py

from os import environ

from agno.agent import Agent
from agno.models.openai.like import OpenAILike

agent = Agent(
    model=OpenAILike(
        id=environ["FOUNDRY_DEPLOYMENT"],
        api_key=environ["FOUNDRY_API_KEY"],
        base_url=environ["FOUNDRY_BASE_URL"],
    ),
    markdown=True,
)
agent.print_response("Share a two-sentence dramatic story.")

The remaining sections document the classic adapter for existing integrations.

Installation

uv pip install -U azure-ai-inference aiohttp agno

Authentication

Navigate to Azure AI Foundry on the Azure Portal and create a service. Then set your environment variables:

export AZURE_API_KEY="your_value_here"
export AZURE_ENDPOINT="your_value_here"  # Of the form https://<your-host-name>.<your-azure-region>.models.ai.azure.com/models
# Optional:
# export AZURE_API_VERSION=***

Example

Use AzureAIFoundry with your Agent:

from agno.agent import Agent
from agno.models.azure import AzureAIFoundry

agent = Agent(
    model=AzureAIFoundry(id="Phi-4"),
    markdown=True
)

# Print the response on the terminal
agent.print_response("Share a 2 sentence horror story.")

Advanced Examples

View more examples here.

Parameters

Parameter	Type	Default	Description
`id`	`str`	`"gpt-4o"`	The id of the model to use
`name`	`str`	`"AzureAIFoundry"`	The name of the model
`provider`	`str`	`"Azure"`	The provider of the model
`temperature`	`Optional[float]`	`None`	Controls sampling randomness; supported values depend on the deployed model
`max_tokens`	`Optional[int]`	`None`	Maximum number of tokens to generate in the response
`frequency_penalty`	`Optional[float]`	`None`	Penalizes new tokens based on their frequency in the text so far (-2.0 to 2.0)
`presence_penalty`	`Optional[float]`	`None`	Penalizes new tokens based on whether they appear in the text so far (-2.0 to 2.0)
`top_p`	`Optional[float]`	`None`	Controls diversity via nucleus sampling (0.0 to 1.0)
`stop`	`Optional[Union[str, List[str]]]`	`None`	Up to 4 sequences where the API will stop generating further tokens
`seed`	`Optional[int]`	`None`	Random seed for deterministic sampling
`model_extras`	`Optional[Dict[str, Any]]`	`None`	Additional model-specific parameters
`strict_output`	`bool`	`True`	Accepted field; does not enable provider schema enforcement in the normal Agent output_schema path
`request_params`	`Optional[Dict[str, Any]]`	`None`	Additional parameters to include in the request
`api_key`	`Optional[str]`	`None`	The API key for Azure AI Foundry (defaults to AZURE_API_KEY env var)
`api_version`	`Optional[str]`	`None`	The API version to use (defaults to AZURE_API_VERSION env var, then `"2024-05-01-preview"`)
`azure_endpoint`	`Optional[str]`	`None`	The Azure endpoint URL (defaults to AZURE_ENDPOINT env var)
`timeout`	`Optional[float]`	`None`	Accepted field, but not forwarded by this adapter
`max_retries`	`Optional[int]`	`None`	Accepted field, but not forwarded by this adapter
`http_client`	`Optional[httpx.Client]`	`None`	Accepted field, but not forwarded by this adapter
`client_params`	`Optional[Dict[str, Any]]`	`None`	Extra arguments forwarded to the classic Azure SDK client; use SDK-supported option names

AzureAIFoundry is a subclass of the Model class and has access to the same params.

Supported Models

Choose from the current model catalog and check the retirement schedule. The adapter accepts a model ID; this does not guarantee availability in your region or account. Several older Llama and Cohere models used in classic examples have retired.

With the classic adapter, Agent(output_schema=...) supplies schema instructions and parses the response afterward. It does not guarantee provider-enforced output, and strict_output does not change that normal path. Check the returned content's type before consuming it. For client configuration, use supported SDK options in client_params or supply the corresponding preconfigured client or async_client.