Support for Azure OpenAI APIs?

[axelsodergard]

I am trying to configure an OpenAI client using Azure OpenAI API, but I keep encountering 404 Not Found and 401 Invalid API Key errors. Despite following the documentation, I am unable to get the client to authenticate and make successful requests.

Could you provide a working example of correctly configuring AsyncOpenAI for Azure OpenAI models? Any guidance on proper base_url structure and authentication would be greatly appreciated.

Thanks in advance!

[kawashishu]

Is this how you set it up? I also got the same error. Moreover, I can't find the model name config parameter for Agent, even the ModelSettings class can't be found.

from azure.identity import DefaultAzureCredential, get_bearer_token_provider
from dotenv import load_dotenv

load_dotenv()

from openai import AsyncAzureOpenAI, AzureOpenAI

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

client = AsyncAzureOpenAI(
    azure_ad_token_provider=token_provider,
    azure_endpoint="https://aoai-eastus2-0001.openai.azure.com/",
    api_version="2024-10-01-preview",
)

set_default_openai_client(client=client)

[coggsflod]

Yes it does:

external_client = AsyncOpenAI(
    api_key="EXTERNAL_API_KEY",
    base_url="https://api.external.com/v1/",
)

spanish_agent = Agent(
    name="Spanish agent",
    instructions="You only speak Spanish.",
    model=OpenAIChatCompletionsModel(
        model="EXTERNAL_MODEL_NAME",
        openai_client=external_client,
    ),
    model_settings=ModelSettings(temperature=0.5),
)

https://openai.github.io/openai-agents-python/models/

[axelsodergard]

I am attempting to get the following approach to work:

from agents import Agent, InputGuardrail,GuardrailFunctionOutput, Runner, AsyncOpenAI, OpenAIChatCompletionsModel, set_default_openai_client
from pydantic import BaseModel

AZURE_OPENAI_API_KEY = "key"
AZURE_OPENAI_RESOURCE_NAME = "resource-group"
AZURE_OPENAI_DEPLOYMENT_NAME = "gpt-4o"
AZURE_OPENAI_API_VERSION = "2024-08-01-preview"

AZURE_OPENAI_ENDPOINT = f"https://{AZURE_OPENAI_RESOURCE_NAME}.openai.azure.com"

custom_client = AsyncOpenAI(
api_key=AZURE_OPENAI_API_KEY,
base_url=f"{AZURE_OPENAI_ENDPOINT}/openai/deployments/{AZURE_OPENAI_DEPLOYMENT_NAME}",
default_headers={"api-key": AZURE_OPENAI_API_KEY},
default_query={"api-version": AZURE_OPENAI_API_VERSION},
)

set_default_openai_client(custom_client, use_for_tracing=False)

class HomeworkOutput(BaseModel):
is_homework: bool
reasoning: str

guardrail_agent = Agent(
name="Guardrail check",
instructions="Check if the user is asking about homework.",
output_type=HomeworkOutput,
model=OpenAIChatCompletionsModel(
model=AZURE_OPENAI_DEPLOYMENT_NAME,
openai_client=custom_client,
)
)

However, this setup does not seem to be functioning as expected. The same API calls work successfully using curl, so the issue may be related to the SDK. Any insights or suggestions on what might be causing this discrepancy would be greatly appreciated.

This is the error code:

Error code: 404 - {'error': {'code': '404', 'message': 'Resource not found'}} OPENAI_API_KEY is not set, skipping trace export

[rm-openai]

I will followup on this, thanks for reporting.

[rm-openai]

I added a PR #110 that I was able to test with a non-Azure provider. I think it should just work with Azure if you swap it out for AsyncAzureOpenAI, would appreciate your help testing with that.

[huqianghui]

It seems it does not work.

My sample as below:

import asyncio
import json
import os

from agents import (
Agent,
AsyncOpenAI,
ModelSettings,
OpenAIChatCompletionsModel,
Runner,
handoff,
)
from agents.extensions.handoff_prompt import prompt_with_handoff_instructions
from dotenv import load_dotenv
from openai import AsyncAzureOpenAI
from pydantic import BaseModel, ValidationError

load_dotenv()

os.environ["OPENAI_API_TYPE"] = "azure"

openai_client = AsyncAzureOpenAI(
azure_endpoint="https://XXX.openai.azure.com/",
azure_deployment="gpt-4o-mini",
api_version="2025-02-01-preview",
api_key="XXXX")

agent = Agent(name="Assistant",
instructions="You are a helpful assistant",
model=OpenAIChatCompletionsModel(
model="gpt-4o-mini",
openai_client=openai_client),
model_settings=ModelSettings(
temperature=0.7,
top_p=1.0))

result = Runner.run_sync(agent, "Write a haiku about recursion in programming.")
print(result.final_output)

I got the below error:

Tracing client error 401: {
"error": {
"message": "Incorrect API key provided: helloworld. You can find your API key at https://platform.openai.com/account/api-keys.",
"type": "invalid_request_error",
"param": null,
"code": "invalid_api_key"
}
}

@rm-openai , any suggestion?

[jurejoy]

Yes it does:
external_client = AsyncOpenAI(
api_key="EXTERNAL_API_KEY",
base_url="https://api.external.com/v1/",
)
spanish_agent = Agent(
name="Spanish agent",
instructions="You only speak Spanish.",
model=OpenAIChatCompletionsModel(
model="EXTERNAL_MODEL_NAME",
openai_client=external_client,
),
model_settings=ModelSettings(temperature=0.5),
)
https://openai.github.io/openai-agents-python/models/

stream mode not work, AttributeError: 'CompletionUsage' object has no attribute 'input_tokens'

pull repo again, just fixed by #112
It works for me.
try to validate by examples/basic/stream_text.py

[rm-openai]

@huqianghui, you'll need to disable tracing. Added a note for that here: https://openai.github.io/openai-agents-python/models/#using-other-llm-providers

[georg-wolflein]

This worked for me (see here):

from openai import AsyncAzureOpenAI
from agents import set_default_openai_client, set_default_openai_api
client = AsyncAzureOpenAI(
    api_key='xxxxxx',
    base_url='https://xxxx.openai.azure.com/',
    api_version="2024-05-01-preview",
)
set_default_openai_client(client, use_for_tracing=False)
set_default_openai_api("chat_completions")

[triple4t]

@axelsodergard

Here's the working code for configuring AsyncOpenAI to use Azure OpenAI models:

import os
from dotenv import load_dotenv
from agents import Agent, InputGuardrail, GuardrailFunctionOutput, Runner, AsyncOpenAI, OpenAIChatCompletionsModel, set_default_openai_client, set_tracing_disabled
from pydantic import BaseModel

load_dotenv()

set_tracing_disabled(disabled=True)

AZURE_OPENAI_API_KEY = os.getenv("AZURE_OPENAI_API_KEY")
AZURE_OPENAI_RESOURCE_NAME = os.getenv("AZURE_OPENAI_RESOURCE_NAME")
AZURE_OPENAI_DEPLOYMENT_NAME = os.getenv("AZURE_OPENAI_DEPLOYMENT_NAME")
AZURE_OPENAI_API_VERSION = os.getenv("AZURE_OPENAI_API_VERSION")

AZURE_OPENAI_ENDPOINT = f"https://{AZURE_OPENAI_RESOURCE_NAME}.openai.azure.com"

custom_client = AsyncOpenAI(
    api_key=AZURE_OPENAI_API_KEY,
    base_url=f"{AZURE_OPENAI_ENDPOINT}/openai/deployments/{AZURE_OPENAI_DEPLOYMENT_NAME}",
    default_headers={"api-key": AZURE_OPENAI_API_KEY},
    default_query={"api-version": AZURE_OPENAI_API_VERSION},
)

set_default_openai_client(custom_client, use_for_tracing=False)

class HomeworkOutput(BaseModel):
    is_homework: bool
    reasoning: str

guardrail_agent = Agent(
    name="Guardrail check",
    instructions="Check if the user is asking about homework.",
    output_type=HomeworkOutput,
    model=OpenAIChatCompletionsModel(
        model=AZURE_OPENAI_DEPLOYMENT_NAME,
        openai_client=custom_client,
    )
)

result = Runner.run_sync(guardrail_agent, "Is this a homework question?")
print(result.final_output)

result = Runner.run_sync(guardrail_agent, "what is 2+2?")
print(result.final_output)

[sanjeev-kumar-4]

@triple4t @axelsodergard The above code works for me as well.

Setting tracing to false in order to use a : is this a bug? When I set tracing to True, the agent returns a response generated by LLM, but post generation, I get a 401 error.

set_default_openai_client(custom_client, use_for_tracing=True)

Error:

ERROR:openai.agents.tracing:Tracing client error 401: { "error": { "message": "Incorrect API key provided: d2dfc141********************d10c. You can find your API key at https://platform.openai.com/account/api-keys.", "type": "invalid_request_error", "param": null, "code": "invalid_api_key" } }

I'll try and dig in. But to get some pointers, it is failing while generating the traces? Is Response API the default mode for Agent SDK?

[Charikshith]

Yes it does:

external_client = AsyncOpenAI(
api_key="EXTERNAL_API_KEY",
base_url="https://api.external.com/v1/",
)

spanish_agent = Agent(
name="Spanish agent",
instructions="You only speak Spanish.",
model=OpenAIChatCompletionsModel(
model="EXTERNAL_MODEL_NAME",
openai_client=external_client,
),
model_settings=ModelSettings(temperature=0.5),
)
https://openai.github.io/openai-agents-python/models/

Thank a lot. This is really helpful . I am able to run Ollama, Deepseek models as well.

[github-actions]

This issue is stale because it has been open for 7 days with no activity.

[github-actions]

This issue was closed because it has been inactive for 3 days since being marked as stale.