Patch AI audio models like you would in a modular synth: sockets, cables, and listening at a node- not a timeline or a one-shot generate button. Pure Data and Max/MSP make wiring feel natural, but they do not fetch weights, check licenses, or hand someone a runnable AI chain. GroovyUI is a local patchbay for that job: search and install models, wire typed sockets (audio, MIDI, text, control), render offline, audition cached outputs (with A/B compare), and share the graph as .groovy.json. Interaction design for AI audio follows artisanal intelligence: the musician stays in the loop. Model Browser (Cmd+K) for search/install/Plan, Patch Generation (Cmd+G) to draft a graph, Tab Creative Agent to grow the next hop as a ghost on the canvas, and drag-select to inspect one node, A/B two, or summarize three or more---Install, Apply, and Tab-accept stay explicit. Trust sits beside the patch: Compliance (Cmd+C) surfaces license and provenance on the active chain, with optional PDF export. Play auditions cached PCM, not a live DAW clock. Generative hops are labeled. We demo Hello GroovyUI, Tab Creative Grow, Podcast Denoise, Prompt Modular Synth, and Isolate-Translate-TTS. Apache-2.0: https://github.com/yyf/GroovyUI.