Skip to content

Questions about DAWN.

Short answers. The documentation has the long ones.

What is DAWN?

A TypeScript SDK and CLI for composing, arranging and rendering music on your own machine, and a Mac app that makes a loop from a prompt and keeps it playing while you edit it. A song is a folder of TypeScript files: tempo, key, sections and tracks. dawn render turns it into a WAV.

The name stands for Digital Audio Warp Navigator.

Is it free?

Yes. There is no paywall, and composing and rendering need no account. The models DAWN uses run on your Mac, and the coding agent is one you already have, such as Claude Code or Codex, on your own subscription. Nothing is metered, so there is nothing to bill.

Is DAWN open source?

Partly. The SDK, the extension kit and the renderer are MIT licensed and public at github.com/hraness/dawn-sdk. That is enough to write a score and render it to audio without the app.

The app, the dawn command and the composer are closed and ship under the DAWN License Agreement. The composer, the part that writes melody, harmony and arrangement, ships compiled.

Why TypeScript, and not SuperCollider, Sonic Pi or Strudel?

Those are great for live coding and synthesis. DAWN is for finished songs: a whole arrangement in a file you can diff, rendered offline to a master you can release, with an agent able to edit the same file.

TypeScript also means dawn check type-checks a song before anything renders, and coding agents already write it well.

Is the music on this site AI-generated?

LOCALHOST was co-written with Claude. Claude drafted each song in the SDK from the producer's ideas, and the producer shaped them through rounds of listening and feedback. It was not generated by dawn album.

What does dawn album do?

It takes a short brief (a title, a seed, a song count and a mix of archetypes) and works through nine cached stages: plan, compose, words, sing, render, mix, master, gate and report. The same brief makes the same album.

Terminal
$ dawn album brief.json --out album

The gate measures each master for loudness, tonal balance, stereo, vocal balance and how many sung words a speech recognizer hears back. A song that fails is made again, up to two times, and the report names any song that still fails. The gates measure a master; they do not decide whether a song is good.

Singing needs a voice installed once with dawn vocals setup. How it works goes through each stage.

Which coding agents work with it?

Any agent that can edit local files and run shell commands. dawn init writes an AGENTS.md with the edit, check and render loop and a CLAUDE.md that points to it, and the project carries the SDK as one typed file, .dawn/sdk/v1.ts. Claude Code and Codex both work.

In the Mac app, comments you leave on notes go to an agent command. The default runs Claude Code, which you install and log in to yourself; you can set another command. The agent guide has the details.

What do I need to run it?

A Mac with Apple Silicon, on macOS 13 or newer. The download is the app with the dawn command inside, and nothing else has to be installed first.

Some commands want more: ffmpeg for .m4a, .opus and video, and uv for vocals and the local models. Each one says so when you run it. Install has the first run.

Can I export WAV, MIDI or stems?

Yes. dawn render writes WAV, MIDI or score JSON, and .m4a or .opus through ffmpeg. --stems <dir> writes each track's stem and the premaster as float WAV in the same render, or 24-bit FLAC with --stem-format flac. --rate 48000 renders at 48 kHz.

The Mac app also saves a song as an MP4 of the piano roll with its audio. There is no export to another DAW's project format; take the MIDI and stems.

Do I own what I make?

Yes. The licence claims no ownership of the songs, scores, recordings or stems you make, and you can publish, perform or sell them.

Optional models that DAWN downloads for you come under their own licences, which DAWN shows before the download. Stable Audio 3 Small, which makes sounds from text, is under the Stability AI Community License; read it before you sell something made with it.

Does it work offline? What leaves my Mac?

Composing, checking and rendering work offline, and the same score renders to the same bytes. The app and the dawn command send no analytics.

The network comes in where you would expect. The app checks dawnav.com for an update once a day, and dawn update asks GitHub. The LOCALHOST songs download from publish.dawnav.com the first time you open each one. Optional models download the first time you ask for them, such as 1.7 GB for the first generated sound, and each file is checked against a pinned hash. A reference you give as a YouTube link is fetched. The agent that answers comments is the one you configured, so Claude Code talks to Anthropic on your account. Signing in and publishing are optional and go to publish.dawnav.com.

The privacy note covers this website.

Does it run on Windows, Linux or an Intel Mac?

Not Windows, and not Intel Macs. On Linux, a script installs the dawn command without the app, but it needs the GitHub CLI logged in to an account with read access to the private repository, so it is not self-serve yet.

How is it different from a DAW?

In a DAW the song lives in a project file tied to that DAW, and you edit it with a mouse. In DAWN the song is TypeScript: it diffs, reviews and merges like code, and an agent edits the same file you do. The Mac app has a piano roll, but the file is the song.

DAWN does not host third-party audio plug-ins or take live MIDI input. When you want those, take the MIDI and stems into your DAW.

How is it different from AI song generators?

A text-to-song generator hands you an audio file. DAWN hands you the score: every note, sound and mix setting is in the project, so you can change one bar and render again.

Models do narrow jobs. A Kokoro voice sings, a whisper model counts the words it hears back, and Stable Audio 3 Small makes a sound when you ask for one. The notes of a loop from a prompt come from a style generator, with no model involved.