Providers - Microsoft Agent Framework

The Microsoft Agent Framework supports multiple AI providers through a unified interface. Each provider offers chat clients that can be used to create agents.

Azure OpenAI

Microsoft’s Azure-hosted OpenAI models with enterprise features.

Installation

pip install agent-framework-core

Azure OpenAI Responses Client

Recommended for new applications:

from agent_framework.azure import AzureOpenAIResponsesClient
from azure.identity import AzureCliCredential
import os

client = AzureOpenAIResponsesClient(
    project_endpoint=os.environ["AZURE_AI_PROJECT_ENDPOINT"],
    deployment_name=os.environ["AZURE_OPENAI_RESPONSES_DEPLOYMENT_NAME"],
    credential=AzureCliCredential()
)

agent = client.as_agent(
    name="Assistant",
    instructions="You are a helpful assistant."
)

response = await agent.run("Hello!")

Azure OpenAI Chat Client

Standard chat completions:

from agent_framework.azure import AzureOpenAIChatClient
from azure.identity import AzureCliCredential

client = AzureOpenAIChatClient(
    azure_endpoint=os.environ["AZURE_OPENAI_ENDPOINT"],
    deployment_name="gpt-4o",
    api_version="2024-10-01-preview",
    credential=AzureCliCredential()
)

agent = client.as_agent(
    instructions="You are helpful.",
    tools=[my_tool]
)

Azure OpenAI Assistants Client

Persistent assistants with threads:

from agent_framework.azure import AzureOpenAIAssistantsClient
from azure.identity import AzureCliCredential

client = AzureOpenAIAssistantsClient(
    azure_endpoint=os.environ["AZURE_OPENAI_ENDPOINT"],
    api_version="2024-05-01-preview",
    credential=AzureCliCredential()
)

agent = client.as_agent(
    model="gpt-4o",
    instructions="You are a helpful assistant.",
    tools=["code_interpreter", "file_search"]  # Built-in tools
)

Authentication Options

from azure.identity import (
    AzureCliCredential,
    DefaultAzureCredential,
    ManagedIdentityCredential,
    ClientSecretCredential
)

# Azure CLI (for local development)
credential = AzureCliCredential()

# Default credential (tries multiple methods)
credential = DefaultAzureCredential()

# Managed identity (for Azure services)
credential = ManagedIdentityCredential()

# Service principal
credential = ClientSecretCredential(
    tenant_id="...",
    client_id="...",
    client_secret="..."
)

client = AzureOpenAIResponsesClient(
    project_endpoint=os.environ["AZURE_AI_PROJECT_ENDPOINT"],
    deployment_name="gpt-4o",
    credential=credential
)

OpenAI

Direct access to OpenAI’s models.

Installation

pip install agent-framework-core

OpenAI Responses Client

from agent_framework.openai import OpenAIResponsesClient
import os

client = OpenAIResponsesClient(
    api_key=os.environ["OPENAI_API_KEY"],
    model="gpt-4o"
)

agent = client.as_agent(
    name="Assistant",
    instructions="You are helpful.",
    tools=[my_tool]
)

response = await agent.run("What's the weather?")

OpenAI Chat Client

from agent_framework.openai import OpenAIChatClient

client = OpenAIChatClient(
    api_key=os.environ["OPENAI_API_KEY"],
    model="gpt-4o-mini"
)

agent = client.as_agent(
    instructions="You are a helpful assistant."
)

OpenAI Assistants Client

from agent_framework.openai import OpenAIAssistantsClient

client = OpenAIAssistantsClient(
    api_key=os.environ["OPENAI_API_KEY"]
)

agent = client.as_agent(
    model="gpt-4o",
    instructions="You are helpful.",
    tools=["code_interpreter"]
)

Anthropic

Claude models from Anthropic.

Installation

pip install agent-framework-anthropic

Anthropic Client

from agent_framework.anthropic import AnthropicClient
import os

client = AnthropicClient(
    api_key=os.environ["ANTHROPIC_API_KEY"],
    model="claude-3-5-sonnet-20241022"
)

agent = client.as_agent(
    name="ClaudeAgent",
    instructions="You are a helpful assistant.",
    tools=[my_tool]
)

response = await agent.run("Hello!")

Streaming Example

agent = AnthropicClient().as_agent(
    name="Assistant",
    instructions="You are helpful."
)

print("Agent: ", end="", flush=True)
async for chunk in agent.run("Tell me a story", stream=True):
    if chunk.text:
        print(chunk.text, end="", flush=True)
print()

Azure AI Foundry

Azure AI Foundry Agent Service (V2) with project-based agents.

Installation

pip install agent-framework-azure-ai

Azure AI Agent Client

from agent_framework.azure import AzureAIAgentClient
from azure.identity import AzureCliCredential
import os

client = AzureAIAgentClient(
    project_endpoint=os.environ["AZURE_AI_PROJECT_ENDPOINT"],
    credential=AzureCliCredential()
)

agent = client.as_agent(
    name="FoundryAgent",
    instructions="You are helpful.",
    model="gpt-4o",
    tools=[my_tool]
)

response = await agent.run("Hello!")

With Hosted MCP Tools

from agent_framework.azure import AzureAIAgentClient
from agent_framework import hosted_mcp_tool

client = AzureAIAgentClient(credential=AzureCliCredential())

# Use hosted MCP server
mcp_tool = hosted_mcp_tool(
    connection_name="my-mcp-server",
    project_endpoint=os.environ["AZURE_AI_PROJECT_ENDPOINT"]
)

agent = client.as_agent(
    instructions="You can use MCP tools.",
    tools=[mcp_tool]
)

AWS Bedrock

AWS-hosted foundation models.

Installation

pip install agent-framework-bedrock

Bedrock Chat Client

from agent_framework.amazon import BedrockChatClient
import os

client = BedrockChatClient(
    region_name="us-east-1",
    model="anthropic.claude-3-5-sonnet-20241022-v2:0"
)

agent = client.as_agent(
    name="BedrockAgent",
    instructions="You are helpful.",
    tools=[my_tool]
)

response = await agent.run("What's the weather?")

Ollama

Local open-source models.

Installation

pip install agent-framework-ollama

Ollama Chat Client

from agent_framework.ollama import OllamaChatClient

client = OllamaChatClient(
    base_url="http://localhost:11434",
    model="llama3.2"
)

agent = client.as_agent(
    name="LocalAgent",
    instructions="You are a helpful assistant.",
    tools=[my_tool]  # Not all models support function calling
)

response = await agent.run("Hello!")

Not all Ollama models support function calling. Models like llama3.2 and qwen2.5 support tools, while others may not.

Provider Comparison

Provider	Package	Use Case	Authentication
Azure OpenAI	`agent-framework-core`	Enterprise deployments with Azure	Azure Identity
OpenAI	`agent-framework-core`	Direct OpenAI API access	API Key
Anthropic	`agent-framework-anthropic`	Claude models	API Key
Azure AI Foundry	`agent-framework-azure-ai`	Managed agents on Azure	Azure Identity
AWS Bedrock	`agent-framework-bedrock`	AWS-hosted models	AWS Credentials
Ollama	`agent-framework-ollama`	Local, open-source models	None (local)

Provider Features

Common Features

All providers support:

Non-streaming and streaming responses
Function/tool calling (model-dependent)
Session management
Middleware
Chat options (temperature, max_tokens, etc.)

Provider-Specific Features

Azure OpenAI & OpenAI

Code Interpreter: Execute Python code
File Search: RAG over uploaded files
Structured Output: JSON schema-based responses
Web Search: Bing search integration (OpenAI only)

from agent_framework.openai import OpenAIResponsesClient

agent = OpenAIResponsesClient().as_agent(
    instructions="You are helpful.",
    tools=["code_interpreter", "file_search", "web_search"]
)

Anthropic

Extended Thinking: Claude’s reasoning mode
Prompt Caching: Reduce costs for repeated prompts

from agent_framework.anthropic import AnthropicClient

client = AnthropicClient(
    model="claude-3-5-sonnet-20241022",
    # Enable prompt caching
    enable_prompt_caching=True
)

Azure AI Foundry

Hosted MCP Tools: Pre-configured MCP integrations
Project-Scoped Agents: Agents tied to Azure AI projects
Memory Integration: Built-in Azure AI Search memory

from agent_framework.azure import AzureAIAgentClient, FoundryMemoryProvider

memory = FoundryMemoryProvider(
    project_endpoint=os.environ["AZURE_AI_PROJECT_ENDPOINT"],
    credential=credential
)

agent = AzureAIAgentClient(credential=credential).as_agent(
    instructions="You have memory.",
    context_providers=[memory]
)

Client Options

Common Options

All clients support these options:

client = ProviderClient(
    # Model selection
    model="model-name",
    
    # API settings
    timeout=30.0,
    max_retries=3,
    
    # Default chat options
    temperature=0.7,
    max_tokens=1000,
    top_p=0.9
)

Runtime Options

Override options per-run:

agent = client.as_agent(instructions="You are helpful.")

response = await agent.run(
    "Be creative!",
    options={
        "temperature": 1.0,
        "max_tokens": 2000,
        "top_p": 0.95
    }
)

Custom Chat Clients

Create your own provider by implementing BaseChatClient:

from agent_framework import BaseChatClient, ChatResponse, Message

class MyCustomClient(BaseChatClient):
    def __init__(self, api_key: str, model: str):
        super().__init__()
        self.api_key = api_key
        self.model = model
    
    async def _inner_get_response(
        self,
        *,
        messages: list[Message],
        options: dict | None = None,
        **kwargs
    ) -> ChatResponse:
        """Call your LLM API here."""
        # Call your API
        response = await self.call_api(messages)
        
        return ChatResponse(
            messages=[Message("assistant", response.text)],
            usage={"total_tokens": response.tokens}
        )
    
    async def _inner_get_streaming_response(
        self,
        *,
        messages: list[Message],
        options: dict | None = None,
        **kwargs
    ):
        """Stream responses from your LLM API."""
        async for chunk in self.stream_api(messages):
            yield ChatResponseUpdate(text=chunk)

# Use your custom client
client = MyCustomClient(api_key="...", model="custom-model")
agent = client.as_agent(instructions="You are helpful.")

Best Practices

Use environment variables for credentials

Store API keys and endpoints in environment variables:

import os

client = OpenAIResponsesClient(
    api_key=os.environ["OPENAI_API_KEY"]
)

Choose the right authentication method

For Azure:

Local development: Use AzureCliCredential
Production: Use ManagedIdentityCredential or DefaultAzureCredential

Set appropriate timeouts

Configure timeouts based on your use case:

client = ProviderClient(
    timeout=60.0,  # Longer for complex tasks
    max_retries=2
)

Verify tool support

Check if your model supports function calling before using tools. Not all models or providers support tools.

Use Azure for enterprise

Azure OpenAI and Azure AI Foundry provide:

Enterprise-grade security
Data residency
SLAs
Managed identity authentication

Provider Selection Guide

Choose Azure OpenAI if you need:

Enterprise security and compliance
Data residency requirements
Azure ecosystem integration
Managed identity authentication

Choose OpenAI if you need:

Latest model releases
Web search capabilities
Simpler setup for development

Choose Anthropic if you need:

Claude’s advanced reasoning
Extended context windows
Strong safety features

Choose Azure AI Foundry if you need:

Managed agent infrastructure
Built-in memory and search
Project-based organization

Choose Bedrock if you need:

AWS ecosystem integration
Multiple model providers (Anthropic, Meta, etc.)
AWS security and compliance

Choose Ollama if you need:

Local, offline operation
Open-source models
No API costs
Full data privacy

Environment Setup

Azure OpenAI

export AZURE_AI_PROJECT_ENDPOINT="https://your-project.cognitiveservices.azure.com"
export AZURE_OPENAI_RESPONSES_DEPLOYMENT_NAME="gpt-4o"
# Authentication via: az login

OpenAI

export OPENAI_API_KEY="sk-..."

Anthropic

export ANTHROPIC_API_KEY="sk-ant-..."

Azure AI Foundry

export AZURE_AI_PROJECT_ENDPOINT="https://your-project.cognitiveservices.azure.com"
# Authentication via: az login

AWS Bedrock

export AWS_REGION="us-east-1"
export AWS_ACCESS_KEY_ID="..."
export AWS_SECRET_ACCESS_KEY="..."

Ollama

# Install Ollama: https://ollama.com
ollama pull llama3.2
ollama serve  # Starts on http://localhost:11434

Agents - Using providers with agents
Tools - Provider-specific tool support
Workflows - Multi-provider workflows

​Azure OpenAI

​Installation

​Azure OpenAI Responses Client

​Azure OpenAI Chat Client

​Azure OpenAI Assistants Client

​Authentication Options

​OpenAI

​Installation

​OpenAI Responses Client

​OpenAI Chat Client

​OpenAI Assistants Client

​Anthropic

​Installation

​Anthropic Client

​Streaming Example

​Azure AI Foundry

​Installation

​Azure AI Agent Client

​With Hosted MCP Tools

​AWS Bedrock

​Installation

​Bedrock Chat Client

​Ollama

​Installation

​Ollama Chat Client

​Provider Comparison

​Provider Features

​Common Features

​Provider-Specific Features

​Azure OpenAI & OpenAI

​Anthropic

​Azure AI Foundry

​Client Options

​Common Options

​Runtime Options

​Custom Chat Clients

​Best Practices

​Provider Selection Guide

​Choose Azure OpenAI if you need:

​Choose OpenAI if you need:

​Choose Anthropic if you need:

​Choose Azure AI Foundry if you need:

​Choose Bedrock if you need:

​Choose Ollama if you need:

​Environment Setup

​Azure OpenAI

​OpenAI

​Anthropic

​Azure AI Foundry

​AWS Bedrock

​Ollama

​Related

Azure OpenAI

Installation

Azure OpenAI Responses Client

Azure OpenAI Chat Client

Azure OpenAI Assistants Client

Authentication Options

OpenAI

Installation

OpenAI Responses Client

OpenAI Chat Client

OpenAI Assistants Client

Anthropic

Installation

Anthropic Client

Streaming Example

Azure AI Foundry

Installation

Azure AI Agent Client

With Hosted MCP Tools

AWS Bedrock

Installation

Bedrock Chat Client

Ollama

Installation

Ollama Chat Client

Provider Comparison

Provider Features

Common Features

Provider-Specific Features

Azure OpenAI & OpenAI

Anthropic

Azure AI Foundry

Client Options

Common Options

Runtime Options

Custom Chat Clients

Best Practices

Provider Selection Guide

Choose Azure OpenAI if you need:

Choose OpenAI if you need:

Choose Anthropic if you need:

Choose Azure AI Foundry if you need:

Choose Bedrock if you need:

Choose Ollama if you need:

Environment Setup

Azure OpenAI

OpenAI

Anthropic

Azure AI Foundry

AWS Bedrock

Ollama

Related