[azure-ai-agentserver-core] Connected MCP tools route returns 404 for hosted LangGraph agent (`/agents/{agent}/tools/resolve`)

[nathan-sumit]

Package Versions

• azure-ai-agentserver-core: 1.0.0b17
• azure-ai-agentserver-langgraph: 1.0.0b17
• azure-ai-projects: 2.0.0b3
• Python: <3.10/3.11/3.12>
• OS: Windows

Description

When using azure-ai-agentserver-langgraph with Foundry connected MCP tools ({"type":"mcp","project_connection_id":"..."}), tool resolution/invocation fails because the connected-tools proxy route returns 404.

The SDK appears to call:

• POST /api/projects/{project}/agents/{agent}/tools/resolve?api-version=2025-11-15-preview
• POST /api/projects/{project}/agents/{agent}/tools/invoke?api-version=2025-11-15-preview

These return 404 in hosted-agent scenarios.

Repro Steps

1. Create a hosted LangGraph agent with use_foundry_tools(...).
2. Include at least one connected MCP tool:
   • {"type":"mcp","project_connection_id":"<connection-name>"}
3. Invoke the agent through project Responses API.
4. Observe failure during connected tool resolution or invocation.

Minimal Agent Snippet

from azure.ai.agentserver.langgraph import from_langgraph
from azure.ai.agentserver.langgraph.tools import use_foundry_tools

foundry_tools = [
    {"type": "code_interpreter"},
    {"type": "mcp", "project_connection_id": "<connection-name>"},
]

# middleware=[use_foundry_tools(foundry_tools)]
# invoke agent via project Responses API

Actual Result

• Connected MCP resolve/invoke path returns 404.
• Agent run fails when trying to use connected MCP tools.

Example symptom (sanitized):

• ResourceNotFoundError / HTTP 404 from connected tools route.
• Route shape: /agents/{agent}/tools/resolve and /agents/{agent}/tools/invoke.

Expected Result

Connected MCP tools should resolve and invoke successfully via SDK without requiring custom monkey patching.

Why this appears to be SDK route/contract issue

1. In 1.0.0b17, FoundryConnectedToolsOperations uses /agents/{agent}/tools/resolve|invoke with api-version=2025-11-15-preview.
2. Project connections are valid and retrievable via AIProjectClient.connections.get/list.
3. Direct MCP JSON-RPC calls to connection targets (tools/list, tools/call) succeed.
4. Failure is isolated to connected-tools proxy route behavior.

Temporary Workaround

Implemented runtime monkey patch in app code:

• Catch 404 from connected-tools resolve/invoke.
• Fallback to direct MCP JSON-RPC (tools/list, tools/call) using connection metadata and auth scope.
• This works, but should not be required in user code.

Impact

• Hosted-agent architectures relying on connected MCP tools cannot use default SDK behavior reliably.
• Forces custom patching and increases operational complexity.

Requested Fix

• Provide a supported fix in azure-ai-agentserver-core so connected MCP tools work without monkey patches.
• If route contract changed, document required endpoint/version and update SDK accordingly.
• Consider built-in fallback or clearer actionable error when route is unavailable.

[github-actions]

Thanks for the feedback! We are routing this to the appropriate team for follow-up. cc @ankitbko @RaviPidaparthi @vangarp.