This project does not have a settled name yet. Naming ideas are welcome, open an issue.

Every sound is synthesized live in the browser from a recipe. There are no audio files anywhere in the product. Pick a category, generate until one fits, hear it in context, export it as a WAV or as a few lines of JavaScript.

Early beta, and very much a work in progress. The core loop works: generate, hear it in context, export. The library is growing, the generators are learning, and a lot of what is here is the first version of itself. Do not read this repo as a finished project. Read it as one you can help build.

I am looking for help, and PRs are welcome. In rough order of how much they would matter:

• Installation and packaging. Today the only way to use a sound is to export it from the running app. An npm package that ships the player and lets a project import a sound is the single biggest missing piece. The player is already standalone and dependency-free.
• Training data. The library grows by curation, and every keep or delete trains the generators. If you have ears and patience this is the most valuable thing anyone can do.
• The generators. New instruments, figures, archetypes and characters, and better training for the ones that exist.
• The front end. The stage animations, the product page, the workbench.
• The docs. Anything in docs/ that is stale or unclear, and the Atlas copy.

CONTRIBUTING.md has the specifics, including how curation contributions get merged. docs/TODO.md is the live work list.

Node 24 (Next 16, React 19, Tailwind 4).

npm install
npm run dev

Open the product at /. At the bottom of the page there is a dev mode link. That is the workbench, and it only exists when you run locally: its routes write curation state to data/pool/*.json on your filesystem, and a production build closes them.

Inside the workbench you can:

• Library: browse everything, sort new keeps into categories by ear, and audit the library against the ear-safety ceilings.
• Variations, Creations, Invent, Wild, Craft, Prospect: generate in batches from each engine, keep what is good, delete what is not. Every verdict on Creations and Invent moves that engine's dice. Keeps from every tab land in the to-sort inbox with no category until you sort them.
• Editor and Import: hand-write a recipe, or paste one back in.
• Calibrate: set the ear-safety ceilings and loudness targets by listening.
• Atlas: everything every generator can say, rendered live from the code and the current training state.
• Dedupe: whole-library similar-pair triage. Identical-only; when unsure, keep both.
• Trash: everything deleted or marked duplicate. Restore lives only here.

Read docs/TRAINING.md before a curation session. It says what each tab's keep and delete actually write.

The short version: a hand-curated library, a handful of generators that draw from and remix it, and a learning layer that shifts each generator's odds every time a curator keeps or deletes what it made. The library and the training data are the same asset.

On provenance: the synth core began as an adaptation of an open source recipe player, and a several open source sound packs seeded the very first library. Everything that generates, learns and curates is original to this project, as are the instruments, figures and characters the generators draw from, and the library has been growing past its seeds ever since. The specific projects and their licenses are in THIRD-PARTY-NOTICES.md, also rendered at /licenses.

The long version lives in docs/:

• GETTING-STARTED.md sets up the workbench and explains each tab.
• HOW-IT-WORKS.md is the explainer, with a glossary.
• ARCHITECTURE.md is what is built right now.
• CLAUDE.md is the rulebook the whole project holds to.

MIT. See LICENSE.