The library

Everything Soundbox makes is saved automatically. There is no "save" button: a generation writes a WAV to disk and a row to the database before you hear it. The library is where you find it again.

What a sound carries

Every clip keeps the settings that made it:

Field	Why it is there
Prompt, negative prompt	What you asked for, verbatim
Model	Which one answered
Seed, steps, CFG	Enough to render it again
Duration, sample rate, size	Facts about the file on disk
Peaks	256 measured points, so every waveform draws without decoding
Source	app, api or mcp (who asked)
Parent	Set on anything derived from another sound

"Reuse this prompt" copies all of it back to the composer, so you can make a variation of an earlier sound.

Finding things

• Search matches prompts and names.
• Categories (SFX, Music, Text to Speech) are separate lists, narrowed by the models in each category, so a music render never appears while you are browsing SFX.
• Tags are yours, free-form, and indexed.
• Favourites are marked with a star.
• Agent-made sounds can be filtered on their own, to see what an agent generated.

Sessions and projects

• A session is a sitting. It is created on your first generation, titled from the prompt, and collects everything made in that run automatically. Use it to find what you made on a given day.
• A project is a folder you put things in yourself. Use it to collect what a track is made of.

A sound inherits its session's project, and can be moved to a project of its own. The library shows whether a sound's project is inherited or set directly.

Editing without destroying

Every tool that changes audio writes a new sound with parent_id pointing at the original. Nothing is overwritten.

A sound's page has a waveform you can drag a selection across, and a toolbar that acts on the selection or the whole clip:

Tool	What it does
Trim	Keep the selection, discard the rest
Normalize	To a LUFS target
Gain	Up or down, in dB
Reverse	Backwards
High-pass, low-pass	Cut below or above a frequency
Pitch	Semitones, without changing the length
Stretch	Length, without changing the pitch

Each Apply produces a new clip and opens it. The original is unchanged.

Stems

If a separator model (HTDemucs) is installed, any sound can be split into stems: vocals, drums, bass, other. The stems are ordinary library sounds with their parent set, so you can search, tag and arrange them like any other.

When you split, choose the stems you want and what happens to the others:

• Scrap them: the stems you did not pick are never written. A split for the drums alone leaves one new sound.
• Keep as one: the rest is summed into a single (rest) sound, the original minus what you took out. The pieces still add back up to the take.

On a sound's page the stems appear beneath it as one synced player. They share the source's length and start together. To arrange them, open the sound in the studio, which splits it and puts one stem on each lane.

Where the files are

~/.soundbox/sounds/<id>.wav

Plain WAV, one per sound, named by id. The library's "reveal" opens your file manager with the file selected. Downloads can be converted to mp3 if ffmpeg is installed. Without it you get the WAV, and the app tells you.

The audio is plain files and the metadata is one SQLite file, so to back up your work, copy ~/.soundbox. Deleting a sound from the library deletes its row and its file, and never a derived clip made from it.

From the API

# Everything, newest first
curl 'http://127.0.0.1:17494/api/sounds?limit=20'

# Only sounds from these models, matching "rain"
curl 'http://127.0.0.1:17494/api/sounds?q=rain&models=audiocpp-stable-audio-3-small-music,audiocpp-ace-step'

# One sound, in full
curl 'http://127.0.0.1:17494/api/sounds/<id>'

Every endpoint is in the API reference.