Pick and mix genres, dial in mood, tempo and vocals, then storyboard an
[mm:ss] timeline. The model composes and renders the whole track — you stay in the producer's chair.
1 Inspiration optional
Tap a ready-made vibe to load a full style instantly — or type your own artist, era, or song and we'll translate it into a filter-safe style. No names are ever sent to the model.
✨ Tap a vibe to load it instantly
Not listed? Describe any artist, era, song, or vibe
2 Genre pick one, or two to mix
AB
3 The vibe
4 The timeline
Drag ⋮⋮ to reorder. Each cue tells the model exactly what happens when.
5 The script
Synced from your choices
Auto-title follows your preset picks. Edit manually anytime.
Composing and rendering audio from scratch…
Full-track generation can take up to a couple of minutes.
Stems are the layers a song is built from. Split this track and you can mute, solo, rebalance, and export your own version.
🎤Vocals — mute for an instrumental
🥁Drums — solo the rhythm
🎵Bass — feel the low end
🎹Other — keys, synths & more
Loading stems into the mixer…
🎙 AI-composed lyrics
Share it with the world
Publishing adds your track to the public gallery and creates a shareable link.
Published 🎉
Recently Generated
No published tracks yet — be the first to compose and publish one.
How this really works
This isn't a search box — it's a tiny record label running on one API call and a database.
You direct. The builder compiles your genre mix, mood, tempo, vocals, and [mm:ss] timeline into a single GhostProducer prompt script — the proven genre + mood + BPM + hero instruments + vocal intent + structure formula.
Lyria composes. The script goes to Google's lyria-3-pro-preview via the Gemini API, which renders real audio (and sometimes lyrics) from scratch.
MongoDB stores it. The audio bytes land in GridFS (so big files stream cleanly); a track document goes into a scoped collection, auto-stamped and isolated to this app.
Atlas Search finds it. Publish a track and it joins a gallery you can full-text search across prompts, titles, genres and lyrics — with a graceful text-match fallback.
The crowd reacts. Likes and plays are atomic $inc counters on the document.
You remix. De-compose splits a finished track into stems via a hosted Demucs model; the layers stream back from GridFS into a Web Audio mixer. Save a remix and it's a tiny document that just points at the parent's stems — no audio duplicated.
Style translation uses a text model to turn "sound like <someone>" into a generic, filter-safe style descriptor — capturing the sound, never the name. (The instant Inspiration vibes are pre-built, so they load with no model call at all.)
Your first song in 5 moves
1
Set a direction. Tap a genre — or two to mix them. Try Reggaeton × EDM. Stuck? Tap a ready-made vibe in Inspiration, or type your own and hit Translate.
2
Shape the vibe. Pick a mood, choose vocals, slide the tempo. The genre seeds smart defaults so you're never staring at a blank page.
3
Storyboard it. Edit the timeline — rename sections, change timestamps, drag to reorder. Watch the script update live.
4
Compose & publish. Hit Compose the track, listen, then publish to share it and climb the gallery.
5
Remix it. De-compose any track into stems — vocals, drums, bass — then mute, solo and rebalance in the mixer. Export your mix as a WAV or share it as its own track.