I have tested the conversation in “Test Chat” mode and in “Test Audio” mode. When Chat mode the Tool Invocation: MCP tool formats the response in one way that successfully passes to the MCP, but in Audio mode it formats it in a different way that fails. How do I get Audio mode to pass the results like Chat mode? https://cdn.discordapp.com/attachments/1471287272396820697/1471287272870641817/e7247d011d4a7c187f42cc1ffddfd60f3291e235_2_180x500.png?ex=698e628d&is=698d110d&hm=e8db8094bf029d8508dbfcbc501416b22f4cb466bfca970c171c981f05d97f79& https://cdn.discordapp.com/attachments/1471287272396820697/1471287273831010490/564cd9abe79b677acad839ceea9a701b57b19b7e_2_414x500.png?ex=698e628d&is=698d110d&hm=8b7a5b05b9b2b5871ae43411f7fd67ff7d1da8e1014817a1d06d4c92b1687784&

This looks like Audio mode is formatting the tool output differently (usually as a string instead of clean JSON), which is why MCP can’t parse it. In Chat mode, the payload stays structured, so it works fine. The fix is to force strict JSON formatting and remove any extra text or wrapping in the Audio pipeline. I can help you standardize both modes so they send the same payload. Are you using any custom processing between Audio input and MCP, or just the default setup? @InstilDevAI

Hi @InstilDevAI

@User https://cdn.discordapp.com/attachments/1471287272396820697/1471555852530876608/e7247d011d4a7c187f42cc1ffddfd60f3291e235.png?ex=698f5caf&is=698e0b2f&hm=71ee44a8ff4b0d82d38cf6f6057fb380f468145204e107b8e644a9048f1da4aa& https://cdn.discordapp.com/attachments/1471287272396820697/1471555853244043368/564cd9abe79b677acad839ceea9a701b57b19b7e_2_414x500.png?ex=698f5caf&is=698e0b2f&hm=712425a35f8daf7e5165f5a886b60a3010ad433361c063c47984fd444aa1df0e&

Can you share your agent id

agent_c429dc7519c619b53ca93fd3c3

Hi @InstilDevAI

call_0767a2efa5878f8014eece9bb42

Ok I check with the team

Thank you so much! Really appreciate it

Hi @InstilDevAI This is a bug with Claude Sonnet 4.5. In the meantime, you can first turn off the structure output of the llm model and see if it helps. If not, try to use a gpt 4.1/5.1 model to avoid the failure before we fix the issue. Generally speaking, Sonnet 4.5 is more expensive and slower than gpt models. Not very recommended. https://cdn.discordapp.com/attachments/1471287272396820697/1471852631428305118/Screenshot_2026-02-13-17-08-04-04_abb9c8060a0a12c5ac89e934e52a2f4f.jpg?ex=69907115&is=698f1f95&hm=59f4e856262049df90aef5fe21f7e32c4913160b1afac6a46a5e56dd9dbcb4f9&

Thanks for the feedback I'll try that

Can you try with 4.1

4.1 works!