Generally, you'll have to build that layer yourself. No toolkits or SDKs that I'm aware of let you control a computer with an arbitrary prompt. That would probably be best relegated to an LLM that offers the user a suggested workflow and lets them confirm it.

> I realize you first need to record the audio

There are implimentations that let you stream audio in realtime, but it is rather finnecky and will require basic knowledge of audio DSP. I never set it up though, ymmv.