Variations of variations
Declare a theme once. Derive everything else with composable transformations of pitch, time, shape, and accent. Variations can name variations as their source — they compose all the way down.
Private beta · open invites every Friday
Finished tracks.
From a conversation.
Zigzag is an AI-assisted audio engineering tool that produces finished tracks from scratch. Describe what you want; an engineer-grade agent composes the parts, dials the mix, masters the master, and gives it back as a track ready to release.
Conversational sessions
Engineer-grade mixes
Mastered to release
Reproducible by design
Featured render
A track. Rendered from text.
Every note, mix decision, and effect on this track lives as text in the project tree. Edit a line, re-render, hear the change.
showcase · master
48 kHz · stereo · 192 kbps mp3master render · incremental on re-run
0:00 / 3:48
tap a bar to seek
rendered just now
What this is
An audio engineering studio with an engineer inside —
not another text-to-audio model.
Zigzag is two things. It’s a body of human-readable text that compiles into a finished, mastered audio track. And it’s a set of tools that make that text easy for an AI to work with — to read, reason about, and rewrite during a working session with you.
Generative models guess.
Zigzag composes.
Suno and friends are giant pre-trained weights — you write a prompt, audio comes out, the "what" of the music lives in the model. Zigzag is a project of human-readable text that deterministically renders to audio. Every musical decision is something you can read, diff, and refine.
The render isn’t the artifact.
The project is.
A finished mix is what a binary is to a compiler — a build output. What you keep, share, and iterate on is the directory: the chord progression, the instruments, the arrangement, the mix. Re-renderable, reproducible, content-addressable.
You don’t prompt it.
You talk to it.
Zigzag’s engineer agent reads your project the way a senior dev reads a codebase — patterns, instruments, FX chains, mix decisions. Ask it to tighten the kick. Ask it to derive a half-time variation. It writes the change; the engine renders it.
It’s not AI making the music.
It’s AI making the program that makes the music.
Every sample of PCM in a render is determined algorithmically — by the patterns you wrote, the modulators you wired, the FX chain and arrangement you composed. The agent’s job isn’t to guess audio; it’s to author the deterministic program whose execution yields the audio. Reproducible. Readable. Refinable line by line.
A DAW for the agent
The whole room,
shrunk into the computer.
Synthesizers defined in code. Routing and effects without limit. Event data like MIDI, with monadic composition and transformation of musical ideas. Completely deterministic: every sample of every render is a function of the project — not a guess from a model.
Every parameter is an expression
LFOs, envelopes, and automation lanes combine with arithmetic — drive a pad with a slow fade modulated by a rhythmic gate. Knobs are formulas, not single values you have to ride.
Stereo-aware effects
Every effect carries a channel side: left, right, or both. A right-side delay yields Haas-style width with no special-casing. Stereo is a property of the chain, not an afterthought.
Sidechain & studio mastering
Classic pump-and-duck compression on any track keyed by any other. A studio-grade multiband splits the signal into bands, compresses each independently, sums them back — house-music glue or transparent mastering on tap.
Voice as an instrument
A spoken clip synthesizes text into audio at render time — inline announcements, click-track callouts, spoken lyrics. Choose a voice from the catalog; swap engines without rewriting the project.
Hot-reload your mix
Watch mode re-renders only what changed and hot-reloads playback. Edit a filter cutoff, save, hear it a second later. The development loop you wish your DAW had.
Send / return busses
Short room verb. Long hall. Per-track sends, ducked by per-bus FX. The classic mixing topology — declared in plain text, summed inside the engine.
Built-in mix analysis
LUFS · true-peak · DR. Per-band stereo width. Onset timing. Anomalies — clip bursts, DC offset, channel asymmetry. The agent reads the report before it touches the mix — and after.
Incremental everything
Every clip and intermediate is keyed by content. Identical inputs become an instant cache hit. A render is instant when nothing changed, partial when little did, and never wastes a sample on what you already heard.
The project tree
Open the project.
Read the music.
A Zigzag project is a directory of plain-text source. Show me the project is now a question with an answer. So is what changed? — every commit is a production diary entry, every mix decision a line you can read out loud.
Source
Plain text
Cache
By content
Output
48 kHz stereo
project · showcase
showcase # the project — versioned, shareable, deterministic
├─ project # tempo · sample rate · channels
├─ arrangement # the timeline · what plays when
├─ chords # the harmonic backbone
├─ instruments/
│ ├─ banjo
│ └─ bodhrán
├─ drums/
│ └─ kit
├─ variations/ # themes deriving themes
│ ├─ bass · half-time
│ └─ pad · chopped & wider
├─ effects/ # reverbs · filters · compression · delays
├─ modulators/ # every parameter is a formula
├─ busses/ # short room · long hall · returns
├─ master · render ← build output, content-addressed
└─ cache · incremental ← keyed by what changed; nothing re-rendered without reason
One theme · four derivations
bass · theme
→
+2 pitch
→
½× time
→
triplet feel
→
bass · fast
Every variation is a derivation, not a copy. Change the source theme; every variation downstream rebuilds with it.
Working session
It’s not a prompt.
It’s a working session with an engineer.
People ask "what prompt did you use?" — but that question doesn’t quite fit. Through conversation with the engineer agent, you and the AI together compose, arrange, mix, and master, at every layer from signal processing to music theory. The result isn’t guessed; it’s constructed, decision by decision.
01 Drums · 808 + hi-hat
you
Hi, I want to hear an 808 with a hi-hat pattern.
zigzag · engineer
Easy starting point — 808 sub on 1 and 3, hat on 8ths, tight and closed. Two bars at 90. Have a listen:
808 + hi-hat 2 bars · 90 bpm · closed hat
0:00 / 0:05
↓ wav
you
Okay, I want it. Can you open the hats on the offbeats?
zigzag · engineer
Done — closed on the downbeat, open on the “and” of each beat. Same kicks, same tempo, just airier hats. Listen:
808 + hi-hat · open 2 bars · 90 bpm · open on the offbeat
0:00 / 0:05
↓ wav
you
Now layer a low bass on the downbeat. Same key.
zigzag · engineer
Added a low F under the kick — hits on 1 and 3 of each bar, voiced so it doesn’t crowd the 808’s sub. Listen:
808 + hi-hat + bass 2 bars · 90 bpm · root F on the downbeat
0:00 / 0:05
↓ wav
02 Full track · end-to-end
you
I want a track in A minor — banjo and bodhrán, a low bass sitting just under the kick, gentle swing on the bridge. About four minutes. Mastered.
zigzag · engineer
Got it. Starting with the harmonic backbone — minor-ish circle of fifths so the bridge doesn’t feel stranded. Bass on a 4-bar loop, voiced low enough to leave room for the kick. Banjo and bodhrán each with their own pre-FX, then a short room and a long hall on shared sends. Targeting −14 LUFS, comfortably below the −1 dBTP ceiling.
you
yes go
↳ arrangement · 16 bars × 4 sections
· bass · cycle source · chord-tones · gentle swing on bridge
· banjo · bodhrán · per-hit expression · scale-locked
· busses · short-room · long-hall · per-track sends
· master · multiband + true-peak limiter
✓ rendered 47 segments · 6 re-rendered (incremental)
↳ master · 3:48 · −14.1 LUFS · true-peak −0.9 dB · DR 9.4
zigzag · engineer
Here it is — 3:48 at −14.1 LUFS, comfortably below the −1 dBTP ceiling. The bridge swings about 18% on the offbeat; happy to tighten if you want it sharper, or pull the bodhrán forward in the mix. Listen:
showcase · master 48 kHz · stereo · 192 kbps mp3
0:00 / 3:48
↓ wav
~340 musicians ahead of you
Make music with an
engineer in the room.
Zigzag is in private beta. New invites every Friday. Tell us a little about what you make and we’ll get you in.
We’ll only email you about the beta and GA launch. No newsletter, no drip, no growth-hack sequences.