Having issues calling MCP servers using Azure Open AI responses api

Microsoft Support Ticket Details

Issue Summary

Azure OpenAI Responses API rejects MCP (Model Context Protocol) tool requests with error: "MCP server url 'mcp.zapier.com' is not allowed. Allowed domains: []."

This was working and stopped

Error Message

HTTP/1.1 400 Bad Request

{

  "error": {

    "message": "MCP server url 'mcp.zapier.com' is not allowed. Allowed domains: [].",

    "type": "invalid_request_error",

    "param": "tools",

    "code": null

  }

}

Environment

• Azure OpenAI Endpoint: https://helioscore.openai.azure.com/openai/v1/responses
• API Version: v1 (Responses API)
• Model: gpt-4.1 (also tested with other models)
• Region: (include your Azure region)
• SDK: OpenAI Python SDK (latest version)

Request Details

We are calling the Responses API with MCP tools as documented in Microsoft's official documentation:

https://dori-uw-1.kuma-moon.com/en-us/azure/ai-foundry/openai/how-to/responses

Request format:

client = OpenAI(

    api_key="<AZURE_API_KEY>",

    base_url="https://helioscore.openai.azure.com/openai/v1/",

)

response = client.responses.create(

    model="gpt-4.1",

    input=[{"role": "user", "content": "Hello"}],

    tools=[

        {

            "type": "mcp",

            "server_label": "zapier",

            "server_url": "https://mcp.zapier.com/api/mcp/mcp",

            "headers": {"Authorization": "Bearer <token>"},

            "require_approval": "never"

        }

    ]

)

Expected Behavior

Per Microsoft documentation, the Responses API should support remote MCP servers. The documentation states:

"You can extend the capabilities of your model by connecting it to tools hosted on remote Model Context Protocol (MCP) servers."

Actual Behavior

The API returns a 400 error stating the MCP server URL is not allowed, with an empty allowed domains list: Allowed domains: []

This blocks all API requests when any MCP tool is included - even for simple queries that wouldn't invoke the MCP tool.

Questions for Microsoft

1. Is there a configuration setting in Azure Portal or AI Foundry to whitelist MCP server domains?
2. Is this a new security feature that requires explicit configuration?
3. If domain whitelisting is required, where is this configured?
4. Is this behavior specific to certain Azure regions or subscription types?

Impact

• MCP tool integration is completely blocked
• This was previously working (unclear when it stopped)
• Affects production application functionality

What We've Tried

1. Using AzureOpenAI client vs standard OpenAI client with base_url - same error
2. Different MCP server URLs (Zapier, custom servers) - same error
3. Different models (gpt-4.1, gpt-4o, gpt-5) - same error
4. Verified MCP tool format matches Microsoft documentation exactly