I’ve thoroughly tested my custom MCP server — it works perfectly with the OpenAI ChatGPT 4.1 model in VAPI. Recently, I switched to the latest OpenAI Realtime model (

gpt-realtime-2025-08-28

) and encountered a problem. I’m using the exact same prompt and the same MCP server setup. In the OpenAI Playground, everything functions correctly — the model calls the MCP tools properly, receives the response, and continues the conversation as expected. However, in VAPI, although the transcript clearly shows that the model successfully calls the correct tool on my MCP server and receives the expected result, the model behaves as if it never received any data — it doesn’t process or acknowledge the tool response at all. Call ID: 019a07a2-7447-7fff-b37a-63c856ac1d76

Suggestion: Double-check that your MCP server's response to Vapi strictly matches the required format and always returns HTTP 200. Even a minor deviation (like line breaks, missing

results

array, or mismatched

toolCallId

) will cause the model to ignore the tool result, especially with newer models. Example of correct response:

json
{
  "results": [
    {
      "toolCallId": "call_123",
      "result": "Your response as single-line string"
    }
  ]
}

Tip: Use a JSON validator and check your call logs for parsing or format errors. Fixing the response format usually resolves this issue with Vapi's latest models. Sources: - [Custom tools troubleshooting](https://docs.vapi.ai/tools/custom-tools-troubleshooting) - [MCP documentation](https://docs.vapi.ai/tools/mcp)

Any updates on my case?

We are having the same issue

We experinced a similar issue. We used a text based model from open AI before, and the tool use was working as expected. We switched to open AI real time model (voice native) and the tool calls work as expected, return 200, and json data is returned but the model does not use it in the call response.