Soundbox

The free, open-source, local alternative to ElevenLabs and Suno.

AI sound effects, music, text-to-speech and voice cloning, generated on your own GPU.
Describe a sound and get a WAV. Write lyrics and get a song. Clone a voice from a
short clip. Arrange it all on a timeline, or let Claude Code, Cursor or any MCP agent
do it for you. Open-weight models, no account, no credits, works offline. MIT licensed.

Download · Hear it · Features · For agents · Models · Docs

Download

The installers carry their own runtime, so nothing else needs installing.
Models download from inside the app the first time you use them. The builds
are not signed yet, so the first launch asks you to confirm:
docs/DESKTOP.md shows where. Linux, or running
from source or in Docker: see Run from source.

Hear it

Everything below was generated by Soundbox from the prompt shown. The samples site has
around 600 more, each with its prompt and model.

Features

• Sound effects. Describe the material, the action and the space and get a
  clip. Duration, seed, steps, CFG and negative prompt are all exposed, and
  every generation is saved with its settings so you can regenerate or vary it.
• Music, with or without lyrics. Instrumentals and loops from a style
  description, or a sung song from a [verse] / [chorus] lyric sheet. Extend
  a track, cover it in a new style, repaint a section, or split it into drums,
  bass, vocals and other.
• Speech and voices. Text-to-speech from a script, voice cloning from a
  short clip or your microphone, and voice conversion. Saved voices live in a
  Voice Lab and are cast onto the lines of a script.
• A studio. A multi-track timeline with per-track effects, recording, MIDI
  input, a piano roll and one-click mixdown. A chat assistant beside it turns
  requests like "mute 1 and 3, add reverb to the vocal" into edits you confirm.
• Built for agents. An MCP server with 21 tools lets Claude Code, Cursor or
  any MCP client generate, speak, arrange and export audio as part of a larger
  task. An Agents page logs every call.
• Licences before you download. Each model's licence and commercial-use
  terms are on its card, and non-commercial weights need an explicit
  confirmation.
• Optional local helpers. A small local language model can rewrite vague
  prompts, write lyrics and run the studio chat, and a small image model can
  draw cover art. Both are off until you turn them on. The assistant can use
  Ollama or LM Studio instead, and artwork an Automatic1111-compatible server.
• Local and offline. One SQLite file and a folder of WAVs under
  ~/.soundbox. CUDA, HIP, Metal, Vulkan or CPU, picked automatically. Nothing
  is uploaded.

Music: a style prompt or a lyric sheet in, a shelf of takes out	Sound effects: every take kept with its prompt, model and seed
Voice Lab: clone a voice, then cast it onto script lines	Models: licence and VRAM shown before you download

For agents

Start Soundbox, then add it to your client.

Claude Code

claude mcp add --transport http soundbox http://127.0.0.1:17494/mcp

Cursor and other MCP clients

{ "mcpServers": { "soundbox": { "type": "http", "url": "http://127.0.0.1:17494/mcp" } } }

Then ask for what you need: "add a door slam and a coin pickup sound to this
game", "voice these three lines as two characters", "write a 20-second
loading-screen loop". The agent gets 21 tools:

Three agent skills ship with Soundbox: soundbox-sfx (prompting),
soundbox-sound-pass (read a game or app, work out every sound it needs,
generate and wire them in) and soundbox-voice-story (write, voice and arrange
dialogue). Install them from the Agents page or with soundbox skills install.
Connection details, the stdio bridge and the full tool reference are in
docs/MCP.md.

Models

Every model runs through audio.cpp, a native runtime, so
there is no Python to set up. Four models are verified end to end and on by
default:

61 more are in the catalog and can be switched on from the Models page:

• Text-to-speech (45): Chatterbox and Chatterbox Turbo, Fish Audio S2 Pro,
  IndexTTS 2 and 2.5, VibeVoice, Higgs Audio v3, Qwen3-TTS, MOSS TTS, VoxCPM 1
  and 2, CosyVoice 3, FireRedTTS3, Maya1, Kokoro, MagpieTTS, Supertonic 3,
  Pocket TTS and more. Several clone from a reference clip.
• Music (6): ACE-Step 1.5 Base, XL Turbo and XL SFT, Stable Audio 3.0
  Medium, HeartMuLa, MiniMax H3.
• Sound effects (2): MiDashengLM-Gen, ControlFoley.
• Stem separation (4): HTDemucs (4- and 6-stem), BS-RoFormer, Mel-Band
  RoFormer.
• Voice conversion (4): RVC, Seed-VC, MioCodec, MeanVC2.

Of these, VibeVoice 1.5B, MiDashengLM-Gen, HTDemucs 6-stem and Chatterbox Turbo
have also been run end to end; the others are marked unverified until someone
reports. The 19 added in October 2026 need audio.cpp 0.9.2 or newer
(soundbox model install audiocpp fetches the latest). Several are non-commercial, and any model
whose licence could not be established is treated as non-commercial. The
table is a summary, not legal advice: read each licence before you ship what a
model made.

Compared with a hosted service

The studio

Pick a builder when you start a project. The Script Builder is for
dialogue: write lines, cast each to a voice, read them, and arrange the takes.
The Music Builder is for tracks and effects: describe one, generate it onto
a lane, then extend, cover, split into stems or layer more. Both share the
timeline, an eight-effect device rack, MIDI and piano roll, recording, the
library and the chat assistant. Export mixes the arrangement down to one new
sound. The full guide, with every shortcut, is docs/STUDIO.md.

HTTP API

Everything the app does is available over a local HTTP API on port 17494.

curl -X POST http://127.0.0.1:17494/api/generate \
  -H 'Content-Type: application/json' \
  -d '{"model":"audiocpp-stable-audio-3-sfx","prompt":"heavy wooden door slamming in a stone corridor","duration_s":3}'

Long jobs return an id with a server-sent event stream. The server binds to
loopback and has no authentication. Every route is listed in
docs/API.md.

Run from source

Requires Node 22.12 or later. No Python, no CUDA toolkit.

git clone https://github.com/Gadurp1/soundbox
cd soundbox
npm install
npm run dev

This starts the server on port 17494 and the UI on http://localhost:5174.
The first run walks you through picking models for your GPU.
docs/INSTALL.md covers the command-line tool, environment
variables and troubleshooting; docs/DESKTOP.md covers
building the desktop app; npm run dev:mock runs the UI with no GPU or server
(see CONTRIBUTING.md).

Docker. For a server or NAS, the server and web UI run headless in one
container with the audio.cpp runtime already inside: CPU by default, CUDA
with an override file.

docker compose up -d --build
docker compose -f docker-compose.yml -f docker-compose.cuda.yml up -d --build

There is no authentication, so keep the port on localhost or behind a reverse
proxy that has some, and set SOUNDBOX_ALLOWED_HOSTS to the name you reach it
by, or /mcp answers 403. Volumes, GPUs and the details are in
docs/INSTALL.md.

Docs

FAQ

Can I sell what I make? It depends on the model. Each model's licence and
commercial terms are on its card in the app and in the table above. Stable
Audio is free under $1M revenue, ACE-Step 1.5 is MIT, and models with an
unestablished licence are marked non-commercial.

Do I need a GPU? No, but CPU generation is slow. A few GB of VRAM covers
the verified models, and each model card shows its footprint.

Does it work on a Mac? There are installers for Apple silicon and Intel,
and audio.cpp runs on Metal. None of the verified models has been confirmed on
a Mac yet, so reports are welcome.

Contributing

Adding a model is a catalog entry, with no new code. Running a listed model and
reporting the result helps just as much, since that is how a model becomes
verified. See CONTRIBUTING.md.

Licence

Soundbox is MIT. The model weights are not: each carries its own
licence, and what you can do with generated audio is governed by the licence of
the model that made it. Soundbox downloads weights at your direction and does
not redistribute them. The example media under site/public/ was generated
with third-party models; its provenance is in ASSETS.md.