MCP tool approval in Workflows: how to approve `mcp_approval_request` events

[ManniArora]

• Package Name: azure-ai-projects
• Package Version: 2.0.0b2
• Operating System:
• Python Version:

Describe the bug

When running an agent workflow that uses an MCP tool requiring approval, the SDK emits
mcp_approval_request events during streaming — but it is unclear how a workflow is expected
to respond to and approve that request.

There does not appear to be a documented or discoverable flow to provide approval for a specific
mcp_approval_request within a workflow execution.

---

Context

I am using the Azure SDK for Python to run an agent workflow that includes an MCP tool.

While streaming workflow execution events, I observe the following sequence:

Workflow created (id: mcp-tool-workflow:2, name: mcp-tool-workflow, version: 2)
Created conversation (id: conv_61a37b3058e19aab00f2sPt5yfJe074bd5kgTntiOeCG9BRhZK)
response.created
response.in_progress
response.output_item.added
workflow_action
response.output_item.added
mcp_list_tools
response.mcp_list_tools.in_progress
response.mcp_list_tools.completed
response.output_item.done
mcp_list_tools
response.output_item.added
mcp_approval_request
response.output_item.done
mcp_approval_request
response.completed

There is no clear way to approve or deny that request programmatically.

---

Question

How is an MCP approval request supposed to be handled in a workflow?

Specifically:

1. How does the SDK expect clients to identify the approval request?
2. What API should be called to approve or deny that request?
3. How is approval bound to a specific workflow execution?

---

[github-actions]

Thanks for the feedback! We are routing this to the appropriate team for follow-up. cc @dargilco @glharper @howieleung @kingernupur @nick863 @trangevi @trrwilson.

[itsdruvaraj]

I couldn't figure this out as well, did you try setting require_approval to 'never' during the creation of agents?

[ManniArora]

My use case requires explicit approval for certain tools (for example, writing to a database). Therefore, I need the ability to set approval at runtime.

[dargilco]

This is supported, but we need to add a sample for this. From @nick863 , who did this with the C# SDK:

"It is done approximately as in the function calling. The workflow is as follows:

1. Send the request to the MCP server as we usually do ("What is my nickname")
2. In the response we need to check for items of type "McpToolCallApprovalRequestItem" (in terms of python it will have a type of "mcp_approval_request")
3. Create the input item, containing the ID of previous item and "approved" set to true or false
4. Create the new stream response with the item from (3) and with the previous response IT set.

We need to keep creating the streaming response, while MCP approval was called."

[ManniArora]

Thanks for the suggestion and the steps, I tried to implement that flow -

What I’m doing right now:

I created a workflow from YAML that contains an agent with a search tool (i.e. MCP tool)
I send the initial request (e.g., "Get all test plans") via 

stream = openai_client.responses.create(
        conversation=conversation_id,
        stream=True,
        input=input_data,
        extra_body=extra_body,
    )

While streaming, I look for items with item.type == "mcp_approval_request", capture the id, and then create input items of the form:

{
                        "type": "mcp_approval_response",
                        "approve": True,
                        "approval_request_id": approval["id"],
                    }

I then start a new streaming response with that approval item as input_data and the previous conversation id set

However, when I run this and then look at traces in Foundry, I’m seeing the following downstream error:

downstream_error: {"error":"Error invoking agent","agent":{"type":"agent_reference","name":"test-search-tool","version":"5"},"details":{"error":{"message":"The following MCP approval requests have approval responses but weren't passed as input: mcpr_0a4d76c683c1605f006980c722b1a881969a54f6a97bafa5f3","type":"invalid_request_error","param":null,"code":null}}}

Here are some logs for your reference -

Workflow created (id: mcp-tool-workflow-agent:1, name: mcp-tool-workflow-agent, version: 1)
Created conversation (id: conv_09589cf95d31861b00aIWWFCE50HrErgbNxV6Jvb7CM2IaNpKL)

📤 Sending initial message to workflow: 'Get all test plans'

📥 Streaming response...
Event 1 type 'response.created'
Event 2 type 'response.in_progress'
Event 3 type 'response.output_item.added'
Event 4 type 'response.output_item.added'
Event 5 type 'response.mcp_list_tools.in_progress'
Event 6 type 'response.mcp_list_tools.completed'
Event 2 type 'response.in_progress'
Event 3 type 'response.output_item.added'
Event 4 type 'response.output_item.added'
Event 5 type 'response.mcp_list_tools.in_progress'
Event 6 type 'response.mcp_list_tools.completed'
Event 3 type 'response.output_item.added'
Event 4 type 'response.output_item.added'
Event 5 type 'response.mcp_list_tools.in_progress'
Event 6 type 'response.mcp_list_tools.completed'
Event 6 type 'response.mcp_list_tools.completed'
Event 7 type 'response.output_item.done'
Event 8 type 'response.output_item.added'
Event 9 type 'response.output_item.done': MCP approval requested for tool 'knowledge_base_retrieve' (server: kb_test_catalog_knowle_8a5nl)
Event 10 type 'response.completed'

📝 Sending approvals for 1 tool(s) (iteration 1)...
  ✅ Approving MCP tool: knowledge_base_retrieve

📥 Streaming response...
Event 1 type 'response.created'
Event 2 type 'response.in_progress'
Event 3 type 'response.failed': FAILED code='None' message='None'