Not a great approach, IMHO. There are no shortcuts or magic bullets. You can't build a reliable agent (voice or otherwise) by simply dumping a bunch of API endpoints via MCP and expecting the LLM to pick which one to use, and when. By doing that you massively increase: - The risk of unreliability. Will the agent use the API endpoint Tools in the right order? Does it know precisely what all the parameters do? You're massively reliant on the quality of Tool and parameter descriptions that you can't change. - The risk of hallucination. While an agent needs information and context, if you give it too much information it increases the risk of hallucination. For Tool usage with write access, that could be disastrous. - Costs. More decisions for the LLM around which API endpoint Tool to use in what order means more token usage and more token spend. Are you happy to write a blank cheque? For vital or high volume tasks it's always better to write traditional deterministic workflows (using no-code workflow builders such as Make).