Jevthoven

# Jevthoven



Prompt-first symbolic-music studio. Describe music in one sentence — *"a wistful 3/4 waltz, soft keys over round bass, light swing"* — and a live TypeSafe Jev model decides every musical unit, sequentially, with a bounded rolling context (recent bars, motif, harmony plan) resent each step. The result is an editable multitrack composition you can play, reshape, and export.

**There is no offline composer, audio model, or hidden song library.** Every note traces to a persisted Jev decision (or a labeled fixture/manual edit).



https://github.com/user-attachments/assets/176c69e4-501e-4b71-8517-957cc692882a



## What it does

- One natural-language prompt → plan (meter, tempo, form, lanes) → instrument choice → harmony → phrase intent → per-lane groove parameters → complete-bar patterns, one Jev `Choice` call each; code renders every pick into notes.
- Piano-roll and arrangement editing: add/move/resize/delete/velocity/quantize, note and track locks, tempo/swing/instrument/volume/mute/transpose — deterministic edits never call the model.
- Plain-English revisions routed through Jev ("make the bass busier in the second half", "set tempo to 80 BPM").
- Region/lane regeneration and whole-piece variations as resumable jobs.
- Pause / resume / cancel / keep-or-discard for every generation job; jobs survive restarts.
- Browser playback via Tone.js, loop regions, complete-bar previews while generating.
- Local SQLite persistence; export project JSON and Standard MIDI Files; validated JSON import.
- **Fixture provider mode** for development — deterministic synthetic decisions, zero API credits.

## Requirements

- Node.js ≥ 22.16
- For live mode: a `TYPESAFE_API_KEY` (https://typesafe.ai). Fixture mode needs nothing.
- Live mode spends credits — roughly one request