A small side project turned into my favorite demo of the Comfy Cloud MCP.
It started as an experiment: modelling with signed distance fields in Blender, with numpy and the OpenVDB module that ships inside Blender's Python. Then my daughter drew 23 cats on index cards, and the experiment had its subject. With Claude Code, every card became a 3D cat, rigged and exported to a Three.js lounge she can play with on her iPad.
The lounge looked finished. It was also completely silent.
So I asked Claude Code for sound. With the Comfy Cloud MCP connected, it called ElevenLabs sound generation once per cat, ACE-Step for a loop of lounge music, then trimmed and levelled everything with ffmpeg and dropped it into the page. I never opened another tab. That's the part I want to show: how little friction there is between "this needs sound" and sound.
The narration in the video is my own voice, cloned with Chatterbox on Comfy Cloud through the same MCP.
If you're building with an agent and need generated audio, images or video, the MCP is the fastest path I've found. Happy to share the workflow files if you're curious.