Setup
Three keys and a few minutes. The only slow part is RunPod pulling the model image the first time.
Before you start
- A Mac (Apple silicon or Intel, macOS 12 or later) or a 64-bit Windows 10 or 11 PC. No GPU is needed.
- A RunPod account with some credit. Ten dollars goes a long way for images; video burns it faster. You can set a spend limit in the RunPod console.
- The keys described in Keys: a RunPod API key, a RunPod S3 API key, and optionally an xAI key.
1. Install
Download the installer for your platform from Download. On macOS open the .dmg and drag CozyGPU to Applications; on Windows run the installer. Both are signed; see Troubleshooting if Gatekeeper or SmartScreen object.
2. Keys
On first launch the setup wizard asks for your keys. Paste the RunPod API key first; the app validates it against RunPod straight away. Then the S3 API key (access key id and secret), then the xAI key if you have one. Each key is checked when you save it and stored encrypted on your machine with the operating system's keychain (Electron safeStorage). They are never sent to CozyGPU.
You can change keys later in Settings. Without the S3 key the app still generates, but custom LoRAs and inputs over 7 MB are unavailable; without the xAI key the Enhance button is off.
3. First job
Pick a model, type a prompt, press Generate. On the first job for a model CozyGPU creates a serverless endpoint in your RunPod account, named cozygpu-<model>, with at most one worker and a ten second idle timeout, then submits a small warm-up job so you can see progress.
The first job on a new machine can take 10 to 30 minutes while RunPod pulls the model (about 10 for image models, up to 30 for video) image to a worker (we measured 11 minutes for Z-Image Turbo and expect about 15 for MiniMax H3). This happens once per endpoint, not once per job; a warm worker answers in seconds.
What ends up in your RunPod account
| Resource | Created when | Cost when idle |
|---|---|---|
Serverless endpoint cozygpu-<model> | First job on that model | Nothing; workers scale to zero |
| Up to three 10 GB network volumes | First custom LoRA or input over 7 MB | RunPod's storage rate per GB per month |
You can delete any of these from the RunPod console at any time; the app recreates what it needs on the next job.
Where your files live
The library (cozygpu.sqlite plus a media folder), uploads, LoRAs and logs are in the app's data folder: ~/Library/Application Support/CozyGPU on macOS and %APPDATA%\CozyGPU on Windows. Back it up like any other folder.