Hey, thanks for the report. This is a known issue: when multiple MCP servers have tools with the same names, Cursor routes all calls to one server instead of the correct one.
Interesting detail: the bug was fixed in 2.3.21, but all fix confirmations were for local STDIO servers. You’re using remote HTTP servers, so it’s possible the fix didn’t cover that transport.
A couple questions:
- If you leave only one server in
mcp.json, do calls to it work correctly? - Can you share a chat example that shows the call going to the wrong server?
Workarounds for now:
- Rename the tools in the server code: if you have access to the
frappe_assistantcode, add a suffix like_stagingor_harvestmasterto the tool names. Confirmed to work. - Start a new chat: routing may be cached per session.
I’ll pass the info about the HTTP transport issue to the team.