This has been reported via email to VAPI support. Has anyone looked at your log? Here I have 2 examples of two members of a squad.(1) assistant has 10 tools and 2 handoff tools. The MCP is a FastAPI-MCP implementation - request.tools show 102Items and (2) assistant has 9 tools and 2 handoff tools. The MCP is a Zapier-MCP - request.tools show 83Items. Prior it making the MCP into 4 smaller toolset the call would fail because the at times the number of tools would be 627Items and even 696Items. The limit is understood to be 128. Claude saw the pattern and concluded that the number is (N_mcp)² + 2 (num of handoffs). As we have reduced N in the 4 MCPs the squad is working correctly. There is no problem calling the tools and get the data from the CRM.

...
 "body": "OpenAI HTTP request",
    "attributes": {
      "category": "model",
      "service": "call",
      "event": "assistant.model.requestAttemptStarted",...

"request": {
        "tools": "[REDACTED-COMPLIANCE]:102items",
        "temperature": 0,
        "model": "gpt-4.1",
        "messages": "[REDACTED-COMPLIANCE]:1items",
        "max_tokens": 250
      },

"request": {
        "tools": "[REDACTED-COMPLIANCE]:83items",
        "temperature": 0,
        "model": "gpt-4.1",
        "messages": "[REDACTED-COMPLIANCE]:15items",
        "max_tokens": 250
      },

Thanks for sharing this, the pattern strongly suggests the MCP layer is duplicating tool schemas during handoffs/request construction, which explains the exponential jump in tool count and the 128 limit failures. Splitting the MCP into smaller toolsets was the right move, but I’d also inspect whether FastAPI-MCP or Zapier-MCP is re-registering tools per session/handoff. I can help you trace this directly. Quick question: does the tool count remain stable between calls, or does it keep growing over time within the same squad? @milynnus

This is taken from Claude. It does not appear to be growing. @User between calls the number remains the same. I used Claude to inject a http request for tools and the number returned is the actual number of tools. Since it handled Zapier the same as my Fastapi-mcp it is probably not a matter of how the mcp had been built. https://cdn.discordapp.com/attachments/1504374897281990677/1504429917499162624/Screenshot_2026-05-14_at_6.26.12_PM.png?ex=6a079dcb&is=6a064c4b&hm=e0676de757efcc519e5c340718cfd7b97cb49b2e5f5379252481c8c906bec270&

checking if this is resolved/solved for you?

There had been no communication from Vapi. What we did was to create multiple mcps with smaller number of tools to avoid hitting the limit.

do you mind booking a quick time with my colleague https://calendly.com/kyle-brunker-vapi/15-min-meeting

Already spoken to him and he add org to receive a fix.

Okay, good to close the ticket then?

Yes