← All postsDescribe the music, get the score: how our AI composer works

Describe the music, get the score: how our AI composer works

· 1 min readaicompositionengineering

Most AI music tools generate audio — a finished waveform you can't open, can't re-orchestrate, can't loop cleanly, and can't ship as adaptive states. We made the opposite bet: Mowjera's AI generates symbolic music — actual notes on a grid.

Why symbolic beats audio for games

How a prompt becomes a score

Type "dark horror score, creeping tension, panicked combat" and the model picks parameters our procedural composer understands: a mood per game state, key, tempo, form length, and an instrumentation palette from the voice registry — sampled grand piano, strings, an acoustic kit, CS-80-style analog brass. The generator then writes the notes under real music-theory constraints, so a bad model day can cost you taste, never correctness.

Bring your own model

It runs on the Claude API for the best drafts — or entirely on your own machine with any OpenAI-compatible server like llama-server. Grammar-constrained output means even an 8B local model cannot return a malformed score.

AI drafts, you polish. That's the honest contract — and because the draft is notes, polishing is actually possible.

Try it: open any score in the composer and hit ✨ Compose with AI.

Share

Hear it instead of reading about it

Every score on the marketplace plays live in your browser — and the ✨ AI composer drafts an adaptive score from one sentence.

Browse the marketplace