gpu.nz

gpu.nz Code

An app with an AI coding agent: the agent runs on your computer, and the model runs on a GPU server that the app rents at gpu.nz.

Download page and all versions

Installation

Download the file for your system from the download page and install it.

  • Windows 10 and 11 (x64): run the installer. The build has no paid signature yet, so Windows may show “Unknown publisher”: click “More info” → “Run anyway”. We are arranging code signing.
  • macOS (Apple Silicon and Intel): open the .dmg and drag the app to Applications. On the first launch open System Settings → Privacy & Security and click “Open Anyway”.
  • Linux: for the AppImage, make the file executable and run it. The .deb and .rpm packages install with your package manager.
chmod +x gpunz-code-*.AppImage
./gpunz-code-*.AppImage
sudo apt install ./gpunz-code-*.deb
sudo dnf install ./gpunz-code-*.rpm

If the download page says “soon” for your system, that build is not published yet.

First launch and sign-in

On the first launch the app asks you to sign in to your gpu.nz account.

  1. Click “Sign in with gpu.nz”: your browser opens.
  2. Confirm the code the app showed. If you are not signed in on the site, sign in; no account yet — sign up.
  3. Return to the app: it already knows you are signed in.

The server is paid from your gpu.nz balance, so the balance must have funds. It is shown in the app, and the button next to it tops it up.

Picking a model

The app offers ready models and lets you connect your own.

Ready models: Qwen 3.8 27B, tested with the agent, and the same model without refusals. For them the app knows which GPU is needed and how long the launch takes.

Your own model: enter a name from Ollama or a link to a GGUF file on Hugging Face and click “Calculate”. The app works out which GPU is needed and shows the hourly price before launch.

devstral-small-2:latest
hf.co/org/repo:Q4_K_M

Models we have not tested with the agent are marked: whether such a model can call tools becomes clear after launch.

The first launch usually takes 5 to 14 minutes: the server downloads the model. You can minimise the window; the app tells you when it is ready.

Server selection modes

Before launching, choose how the app picks a server.

ModeWhat the app picks
OptimalA reliable host, a fast connection, no expensive traffic. The mode for most tasks.
FasterSpeed comes first: the app looks for the server that answers fastest.
CheaperPrice comes first: the app looks for the most economical server.

If a host fails, the app takes another server and the conversation continues. Money for a failed launch returns to your balance.

Idle shutdown

Next to the launch button, choose when to shut the server down when idle: after 15, 30 or 60 minutes, or never.

The server is also deleted when you quit the app; the app asks first.

While the server runs it is billed by the minute. If you chose “never”, delete the server yourself when you no longer need it.

The agent: project, commands, permissions

The agent works in your project folder on your computer: it reads and edits files, runs commands and tests.

Open your project folder in the app and describe the task in plain text.

Before destructive commands, such as deleting files, the agent asks for permission: “Deny”, “Allow once” or “Always allow”. Read the command before you allow it.

Your project code is not uploaded anywhere. Only requests to the model go to your rented server.

MCP, skills and rules

gpu.nz Code is built on OpenCode, so MCP servers, skills and rules work the way they do in OpenCode.

File or folderWhat for
opencode.jsonproject settings and MCP server connections
AGENTS.mdrules for the agent, in the project root
.opencodea project folder with skills and commands

Configuration details are in the OpenCode documentation: opencode.ai/docs

Updates

On Windows and Linux updates install themselves. On macOS the app tells you that a new version is out: download it from the download page. The app installs an update only when its signature is valid.

Changelog

Error reports and feedback

The app sends us anonymous reports about its own errors: the error text, the app version and the operating system. Your account, project files and requests to the model are not sent.

One checkbox turns the reports off: Help → “Send error reports”.

“Help and feedback” sends us a message right from the app: report a bug, suggest an idea or leave feedback.

What is stored and why is described in the personal data processing policy. Privacy policy

Where the data lives

This is where the app keeps its settings and data:

SystemFolder
Windows%APPDATA%\nz.gpu.code
macOS~/Library/Application Support/nz.gpu.code
Linux~/.config/nz.gpu.code

Uninstalling

  • Windows: Settings → Apps → gpu.nz Code → Uninstall.
  • macOS: drag the app from Applications to the Trash.
  • Linux: remove the package with your package manager, or delete the AppImage file.

A rented server is deleted from the app or in the console at gpu.nz. While it exists, it is billed.

To remove the settings too, delete the folder from “Where the data lives”.

If something goes wrong

The launch takes long

On the first launch the server downloads the model, which usually takes 5 to 14 minutes. You can minimise the window; the app tells you when it is ready. If the launch fails, the money for it returns to your balance.

“No free server”

Pick another selection mode or try again in a few minutes: free servers appear and disappear.

The sign-in code does not work

The code is confirmed in the browser at gpu.nz, not typed into the app. Open the sign-in window again and compare the code in the browser with the one the app shows.

Windows will not run the installer

The build has no paid signature yet: click “More info” → “Run anyway”.

Did not find the answer — write from the app (“Help and feedback”) or to support on the site. Contact support