Setup

Install Cadence

Download the installer for your system, or run from source. Either way, full music generation needs the local backends set up - the installer ships the app and engine, not the multi-gigabyte models.

Download for Windows (.msi) Release notes

Pre-release · v0.2.100 · 24 MB · unsigned, so SmartScreen warns (More info → Run anyway). The app starts and stops its own engine.

macOS and Linux under testing
Cadence_0.2.100_aarch64.dmg (macOS, Apple silicon) and Cadence_0.2.100_amd64.deb (Linux, Debian and Ubuntu). Both are built and smoke-tested on every release, so the app and its engine start and run. Making actual music on them is not verified yet: the models want an NVIDIA GPU, so Linux needs CUDA, and Apple silicon support is unconfirmed. Windows is the tested path. Intel Macs are not covered by this build.

What works today: the full loop - record and train a voice, generate a song from a prompt (or paste your own lyrics) in that voice, manage the provider, API keys, and GPU status in Settings, and browse, play, and export past tracks in the Library. Voices can be renamed and deleted, autotune and smoothing shape how a converted vocal sits, and every song lands in its own dated folder. Dark mode and the Windows installer have shipped, and macOS and Linux builds now ship with every release while they are under testing. See the roadmap.

The short version

Quick start

Download the .msi and install it. Windows SmartScreen warns because it's unsigned - click More info → Run anyway.
Set up the models once with setup-backends.ps1 (needs git and uv) - full command in the next section.
Open Cadence. Wait for the dot in the sidebar to turn green: Engine online.
Type a prompt and click Generate. The first song downloads the models (~10 GB), then plays.

System requirements

Can your PC run it?

Component	Minimum	Recommended
Windows	10, 64-bit	11
GPU	NVIDIA with 8 GB VRAM (CUDA)	RTX 3080-class or newer, 10 GB+ VRAM - what Cadence is developed and tested on
RAM	16 GB	32 GB
Disk	30 GB free	50 GB+ on an SSD - the AI backends and model weights take ~20 GB, plus your tracks
Internet	Broadband for the one-time setup (~10 GB of model weights). Generation runs offline after that
Microphone	Any, for voice training	A decent mic in a quiet room - cleaner takes train a better voice

No compatible GPU? Everything still runs on the CPU - expect minutes per song instead of seconds, and slow voice training. The one-time model setup also needs git and uv installed.

Step by step

Getting started

Download and install

Grab the installer and run it. Because it isn't code-signed yet, Windows SmartScreen shows a blue warning - click More info, then Run anyway. It installs to C:\Program Files\Cadence.

Download for Windows (.msi)

Set up the AI models (one time)

The installer ships the app and its engine, but not the ~10 GB music and voice models. This script installs them into Music\Cadence\vendor, where Cadence looks. (The app can also show you this checklist itself - click See what's missing on the Generate screen.) You need git and uv on your PATH. Open PowerShell and run:

PowerShell

git clone https://github.com/JoeMighty/Cadence.git
powershell -ExecutionPolicy Bypass -File Cadence\scripts\setup-backends.ps1

Already built the backends in a source checkout? Move them across instantly instead of rebuilding: setup-backends.ps1 -MoveFrom .\engine\vendor - the checkout keeps working via junctions it leaves behind.

Open Cadence and check it's live

Launch Cadence from the Start menu. It starts and stops its own engine - nothing to run in a terminal. Watch the dot at the bottom of the sidebar: it turns green “Engine online” within about 10 to 20 seconds.

Make your first song

Type a description like “a calm lo-fi piano loop”, leave the voice on Instrumental, and click Generate. The first run shows “Starting music model server” for ~20 seconds while the model loads, then downloads the model weights (~10 GB, first time only) and plays the result.

Optional - voice, lyrics, and better wording

Train your own voice on the Voice tab, then pick it instead of Instrumental. To write your own lyrics, open + Advanced under the prompt and use the Lyrics box (with an Insert example template) - shape it with [Verse], [Chorus], and [Bridge] tags. For sharper multilingual lyrics, add a Claude key in Settings, or install a local model:

PowerShell · optional

ollama pull qwen3.5:9b

For developers

Build from source

Prefer to run the latest code, or hack on Cadence? Skip the installer and run it from the repository. You'll need these on your PATH first:

Component	Version	Notes
Windows	10 or 11, 64-bit	The tested platform. macOS (Apple silicon) and Linux (Debian, Ubuntu) also ship, under testing
NVIDIA GPU	8 GB VRAM or more	Recommended for generation. CPU works, slowly
Node.js	22 or newer	Frontend and Tauri CLI
Rust	stable, MSVC	Install via rustup - pulls in the MSVC target
uv	latest	Manages the engine's Python environment
Git	any recent
Ollama	optional	Local lyrics and prompt understanding, no API key

From source

Run it in four steps

Clone the repository

Everything lives in one repository - the desktop app and the engine.

PowerShell

git clone https://github.com/JoeMighty/Cadence.git
cd Cadence

Start the engine

uv creates the Python environment and installs dependencies on first run. Leave this terminal open.

PowerShell

cd engine
uv sync
uv run uvicorn app.main:app --port 8000

Start the app

In a second terminal. The first run compiles the Rust shell, which takes a few minutes - later runs are fast.

PowerShell

cd app
npm install
npm run tauri dev

Verify

A Cadence window opens. The badge under the title should read Engine online - that means the shell and the engine are talking. Optionally, add a local model for lyrics later:

PowerShell · optional

ollama pull qwen3.5:9b

If something breaks

Quick checks

Generation fails: “the music model isn't installed”

The AI models aren't in place yet. Click See what's missing on the Generate screen for a live checklist, or run the one-time setup-backends.ps1 from step 2 - it installs the models into Music\Cadence\vendor, where Cadence looks. Settings → Storage shows the exact folder the engine is using, and lets you point it somewhere else.

The sidebar dot says “Engine offline”

The app starts its own engine, so give it 10 to 20 seconds after launch. If it stays offline, fully quit Cadence and reopen it. Still stuck? Open Task Manager → Details and end any lingering cadence-engine.exe, then relaunch. (Running from source instead? Make sure your uv run uvicorn terminal is up and nothing else owns port 8000.)

Windows blocked the installer (SmartScreen)

The installer isn't code-signed yet, so Windows shows a blue “Windows protected your PC” screen. Click More info, then Run anyway. Code signing is on the list.

How do I update to a newer version?

Open Settings → About and click Check for updates. If a newer release exists it shows a Download button; install it over the top - your library and trained voices are kept separately and carry over.

How do I check my GPU is visible?

Run nvidia-smi in a terminal. It should list your GPU and driver version. If the command doesn't exist, install the current NVIDIA driver first. Without a CUDA GPU, generation falls back to CPU and takes minutes instead of seconds.

Building from source: npm run tauri dev won't compile

Confirm Rust has the MSVC toolchain: run rustup show and look for stable-x86_64-pc-windows-msvc. Visual Studio C++ Build Tools must be present - the rustup installer offers them if they're missing.

Still stuck? Head to Support.