Home-Level Agent MCP Calls Take 10+ Minutes (Cold Start Issue)

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Summary
MCP tool calls from the home-level agent (global Cursor agent) take 10-12 minutes to return results, while the exact same MCP calls from a project-side agent return in 1-3 seconds.

Environment
Cursor Version: [Check Help → About]
OS: Windows 10/11
MCP Server: Custom Python/FastMCP stdio server
Python: 3.10+ in venv

Steps to Reproduce

Steps to Reproduce
Configure a custom MCP server in %USERPROFILE%.cursor\mcp.json:

{
“mcpServers”: {
“MyMCPServer”: {
“type”: “stdio”,
“command”: “C:\path\to\.venv\Scripts\python.exe”,
“args”: [“C:\path\to\mcp_launcher.py”]
}
}
}
Open a project-side agent (in a workspace with the MCP server’s project open)

Call an MCP tool → Returns in 1-3 seconds ✓

Open the home-level agent (Cmd+Shift+I or global agent outside any project)

Call the same MCP tool → Takes 10-12 minutes ✗

Expected Behavior

MCP tool calls should complete in similar time regardless of which agent context (home vs project) initiates the call.

Operating System

Windows 10/11

Version Information

Version: 2.x.x
VSCode Version: 1.105.1
Commit: abc123…

For AI issues: which model did you use?

Claude-4.5-opus-high-thinking
Claude-4.7-opus-high-thinking

Additional Information

MCP tool calls should complete in similar time regardless of which agent context (home vs project) initiates the call.

Evidence
Test 1: Project-side agent (fast)
User: “Clear cache on 384 environment”
Agent: [Calls MCP tool]
Result: HTTP 200, Success - returned in ~2 seconds
Test 2: Home-level agent (slow)
User: “What’s the WF instance status?”
Agent: [Calls MCP tool]
Result: Same HTTP 200, Success - but took 10-12 minutes to display
Subprocess Verification
Running the CLI directly shows it completes in <15 seconds:

.venv\Scripts\python.exe tririga_api.py --env 384 admin-workflow-agent workflow-instance

Completes in ~11 seconds with result

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor
No - Cursor works, but with this issue

Hi Avinash, thanks for the detailed report and the comparison between project-side and home-level agent contexts.

MCP tool calls are handled entirely by the local extension host (not our servers), so the difference you’re seeing between home-level and project-side agents likely comes down to how the MCP server connection is initialized in each context. To dig into this, I’d need two things:

  1. Your actual Cursor version — go to Help > About and share the full version string (the “2.x.x / abc123” in your post looks like a placeholder)

  2. MCP output logs from both contexts:

    • Open View > Output, then select MCP Logs (or MCP: MyMCPServer) from the dropdown

    • First, reproduce the fast call from a project-side agent and copy the output

    • Then reproduce the slow call from the home-level agent and copy that output too

The logs will show where the delay is happening — whether it’s during server startup, transport connection, or the tool execution itself.

One quick thing to try in the meantime: after opening the home-level agent, check if the MCP server shows as connected in the MCP panel before you send the tool call. If it’s not connected yet, the first call may be waiting for initialization. A second call after the first one completes would confirm whether this is purely a cold-start issue.