tool_resources.mcp is not saved on agent and thread level

[jozkee]

I noticed tool_resources can be passed to multiple APIs, however is not remembered when used in agent and thread level, I'm not sue if this is only the case for MCP.

curl --request POST --url "https://me.services.ai.azure.com/api/projects/my-project/assistants?api-version=2025-05-15-preview" \
  -H "Authorization: Bearer $MY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o-mini",
    "name": "my-mcp-agent",
    "instructions": "You are a helpful agent that can use MCP tools to assist users. Use the available MCP tools to answer questions and perform tasks.",
    "tools": [
      {
        "type": "mcp",
        "server_label": "deepwiki",
        "server_url": "https://mcp.deepwiki.com/mcp"
      }
    ],
    "tool_resources": {
      "mcp": [
        {
          "server_label": "deepwiki",
          "headers": {},
          "require_approval": "never"
        }
      ]
    }
  }'

Response, check the tool_resources property:

{
  "id": "asst_123",
  "object": "assistant",
  "created_at": 1753919431,
  "name": "my-mcp-agent",
  "description": null,
  "model": "gpt-4o-mini",
  "instructions": "You are a helpful agent that can use MCP tools to assist users. Use the available MCP tools to answer questions and perform tasks.",
  "tools": [
    {
      "type": "mcp",
      "server_label": "deepwiki",
      "server_url": "https://mcp.deepwiki.com/mcp",
      "allowed_tools": null
    }
  ],
  "top_p": 1.0,
  "temperature": 1.0,
  "tool_resources": {},
  "metadata": {},
  "response_format": "auto"
}

Again, this also happens with threads. Looks like tool_resources is only honored in run level.

[amynic]

HI @jozkee - could you share more details on the services/technologies you are using so the community can provide better guidance? Looks like possibly Azure AI Agent Service (project in Foundry) via REST API calling an MCP server you have created and hosted?

Also worth taking a look at this question from the forum a while back about Agent, Run and Thread level access to tools using the Azure AI Agent Service https://github.com/orgs/azure-ai-foundry/discussions/98

[jozkee]

Yes, in a project in Azure AI foundry, I'm using the assistants API.

I was following https://learn.microsoft.com/azure/ai-foundry/agents/how-to/tools/model-context-protocol-samples?pivots=rest and I was trying to make use of tool_resources to tell my mcp server to never require approvals. Then I noticed you can also pass tool_resources to assistants and threads endpoints.

https://github.com/orgs/azure-ai-foundry/discussions/98

I saw that and it explains why tool_resources are not saved at thread level but doesn't explain why it doesn't at agent level. You can also notice that the assistants response confirms that the tool is saved, but not the tool resources.

[amynic]

Based on Microsoft Learn documentation and related GitHub discussions (particularly discussion #98), this behaviour is currently by design in Azure AI Foundry Agent Service. Here's why:

Current Behaviour (as documented)

• Agent Level: tool_resources at the agent level are not persisted in the response, even though they can be provided during creation (as tools)

• Thread Level: Thread-level tool configuration is currently non-functional, even though the API signature accepts tool_resources. Threads are passive containers for messages and don't actively invoke tools.

• Run Level: tool_resources work correctly at the run level, as documented in the Microsoft Learn samples.

Technical Explanation from Microsoft Learn
According to the official documentation:

• Agent-level tools are defined in the tools array and persist across runs
• Thread objects are passive containers for messages and context
• Run-level tool_resources are where MCP server configurations (headers, approval settings) are actively processed

Why This Matters - the current design separates concerns:

• Tools (defined at agent level) specify WHAT tools are available
• Tool Resources (defined at run level) specify HOW those tools should be configured and authenticated

What This Means for Your Implementation

• Define MCP tools at the agent level using the tools array
• Configure MCP settings (headers, approval mode) at the run level using tool_resources
• Don't rely on agent or thread level tool_resources persistence for MCP configurations

adding @lindazqli to confirm this is suggested for MCP integration

[amynic]

Our agents team have also recommended taking a look at this C# sample: https://github.com/Azure/azure-sdk-for-net/blob/main/sdk/ai/Azure.AI.Agents.Persistent/samples/Sample26_PersistentAgents_MCP.md

and one for streaming scenarios: https://github.com/Azure/azure-sdk-for-net/blob/main/sdk/ai/Azure.AI.Agents.Persistent/samples/Sample27_PersistentAgents_MCP_Streaming.md

This uses an MCP tool definition in Azure.AI.Agents.Persistent