Back to Things RSS

Auto Music Creation to Distribution Pipeline

Five-phase autonomous music pipeline
Operator · Sonnet
Spotter · Gemini Flash
Supervisor · Opus
Audio Analyst · Gemini Pro
Human (2 steps only)
Fully Autonomous
Why this is interesting: this workflow was messier than customer support or back-office operations: more tools, more judgment, more desktop coordination, more failure points. If you can close a loop like this, simpler operational jobs stop looking very exotic.
What is true: this fully worked. It took roughly 60 to 100 hours to program and around sixty-plus dollars of compute for the published song. The big unlocks were MCP plus switching Hammerspoon configurations before opening things so the CUA landed in the right places, Gemini feedback on rendered audio, Sonible-assisted mixing inside Ableton, and Openwork acting as a centralized worker that could absorb skills, MCPs, and model switching. If something on this page says CUA, read it as some mix of CUA, MCPs, and Hammerspoon coordination.
What changed my mind: after I actually posted one, it ended up with the most plays on the Spotify account, which was also the moment it became clear this probably should not exist as a product. This diagram is also compressed; the real workflow involved even more switching, AI tools, and automation layers than shown here.
Phase 01
Ingestion & Generation
You
Human Input
Write Lyrics
Lyrics in Apple Notes. The original creative input in the pipeline.
Operator · MCP
Pull Lyrics from Apple Notes
Apple Notes MCP reads lyrics into pipeline.
Operator · 5.4
Clean and format lyrics
Uses GPT-5.4 to fix spelling, clean formatting, and apply general song-structure guidance before generation. It was not rewriting the song, though it would sometimes add something like a repeated hook or a quieter bridge.
Operator · CUA + MCP
Generate on Suno & Udio
Uploads the cleaned lyric prompt and structure cues. Suno handled vocals, Udio handled instrumentation, with batch generation across both.
Operator · CUA + MCP
Stem Splitting
Auto-splits generated tracks into isolated stems for production, including the vocal stem that gets routed into the voice-replacement loop.
Operator · Kit.ai + MCP
Voice Replacement Loop
Uploads the extracted vocal stem into Kit.ai, swaps in a clone voice, downloads the result, and repeats passes until the score is good enough to send forward.
Spotter
Monitors Suno and Udio UI navigation.
Supervisor
Recovers failed uploads.
Phase 02
Evaluation & Curation
Operator · CUA + MCP
Upload & Compare
Compares batch against an existing catalog for tonal consistency.
Supervisor · Opus
Visual and structural pass
Looks for pattern fit, arrangement structure, and obvious visual or textual consistency signals. It could spot alignment, but it was not really hearing the music.
Gemini Pro · Audio
Sonic Analysis
This is where the real audio judgment happened: multi-pass evaluation of frequency balance, dynamics, artifacts, and overall mix quality from rendered audio.
Operator · CUA + MCP
Auto-Regeneration
No contenders pass, so it regenerates on Suno and Udio with adjusted parameters.
Phase 03
Production & Mastering
Operator · Sonnet
Build Ableton Project
Sets up the Live session, loads stems into pre-made tracks, configures macros and routing, uses AI-assisted Smart EQ and Sonible plugin scaffolding, and relies on MCP plus switching Hammerspoon configurations so the CUA can open things into the right places before each pass.
Operator · Kit.ai + MCP
Bring the cloned vocal back in
The approved Kit.ai vocal passes get brought back into the Ableton session so the production loop is working on the replaced voice, not the original singer.
Self-Correcting Audio Loop · N Passes
Operator · Sonnet
Export Audio from Ableton
Clicks export, navigates the local file system, and saves a real .wav to local disk.
Operator · File I/O + MCP
Upload .wav to Gemini
Sends the rendered audio file to Gemini Pro for analysis.
Gemini Pro · Acoustic Intel
Analyze & Issue Directive
Evaluates transients, tonal balance, frequencies, and artifacts from the rendered audio file and outputs strict text feedback that is actually usable on the next pass.
Directive Example
"Macro-transient attack too sharp. Pull back Macro 2. Reduce harshness at 4kHz."
Supervisor · Opus
Map Feedback to UI Actions
Translates Gemini feedback into specific Ableton macro coordinates and values for the operator, while the worker can swap models and skills based on the task it is handling.
Operator · Sonnet
Implement Changes in Ableton
Adjusts EQ, levels, effects, and arrangement using the Sonible AI mixing bundle plus tight snap-grid discipline in Ableton. Then renders again and loops.
Spotter · Flash 3.0
Real-time visual watchdog. Verifies the operator hits the correct knobs and UI targets.
Supervisor · Opus
If the spotter flags a miss, it pauses the operator, recalculates the UI sequence, and restarts safely.
Operator · CUA + MCP
Upload to LANDR
When Gemini outputs approved, the operator pushes the final .wav to LANDR and gets multiple master balances back.
Operator · CUA + MCP
Download Masters from LANDR
Downloads all master options to local disk and sends each one to Gemini for evaluation.
Gemini Pro · Audio
Evaluate Master Options
Listens to all master balances and selects the best candidate for approval.
You
Approval Queue
Approve or Reject Master
Listen. Approve to move on, or reject with notes and send it back into the loop.
Phase 04
Distribution & Publishing
Operator · CUA + MCP
DistroKid Upload Flow
Logs into DistroKid, clicks upload, and enters the release metadata and credits.
Operator · CUA + MCP
Metadata & Payment
Sets post-release date, storytelling fields, and authorizes the release fee.
Artwork Generation · ~32 Canva Renders Avg
AI Agents · #3–6
Generate Album Artwork
Canva AI generates covers. A judging layer narrows the set to the best match.
Operator · CUA + MCP · Canva
Crop 3000×3000 & Upload
Downloads the selected art, resizes it in Canva, and uploads it into the release flow.
Published
Hits Publish
Live on streaming platforms.
Spotter
Monitors DistroKid form fills and Canva navigation.
Supervisor
Validates metadata before publish.
Phase 05
Marketing & Outreach
Operator · CUA + MCP
Search Curator Database
Crawls playlist and curator sources for genre and vibe matches.
AI Agent · Vibe Match
Curator Matching
Matches the track to best-fit curators by playlist vibe and genre overlap.
AI Agent · Outreach
Generate Pitch Emails
Drafts personalized outreach for each target instead of generic release spam.
Operator · CUA + MCP
Send Outreach
Pushes the actual emails out and tracks responses.
Spotter
Validates send targets.
Supervisor
Reviews pitch quality before send.
2
Human Touchpoints
60-100h
Programming Time
8+
AI Agents
$60+
Compute For Published Song
~32
Artwork Renders Avg
Openwork
Centralized Worker Layer