01
No inbound ports
The agent dials out to Magpy over an outbound WebSocket and holds it open. Nothing on your machine is exposed to the internet, and no firewall changes are needed.
Local GPU agent · Linux & Windows
Run your AI on your own GPU
The Local GPU agent links a machine you own to your Magpy workspace, so inference runs on your hardware and your prompts never leave it. One command on Linux, an installer on Windows.
Install
Pick your platform.
Install on the machine with the Nvidia GPU. Both paths end the same way: an agent running in the background, enrolled to your workspace, serving inference from Ollama on that box.
🐧 Linux — one command
Debian / Ubuntu, x86-64 or arm64. Installs Ollama if it's missing, pulls the starter models, installs the agent, enrolls the device, and leaves it running as a service.
Terminal — Linux (Debian / Ubuntu)
curl -fsSL https://www.magpy.co/download/install.sh | bashPrefer to review before running? Read the install script, or do a no-op pass that prints every action without executing it: curl -fsSL https://www.magpy.co/download/install.sh | OR_DRY_RUN=1 bash
🪟 Windows — installer
Windows 10 or 11, 64-bit. Registers the agent as a Windows service and adds a tray icon you enroll and manage it from.
Installer — Windows 10 / 11 (64-bit)
Install Ollama first. The Windows installer does not install it — it checks for it and sends you to ollama.com/download/windows if it's missing. Install Ollama, then run this installer.
Windows will warn you that the publisher is unknown. The installer isn't code-signed yet, so SmartScreen shows a blue “Windows protected your PC” screen. Choose More info → Run anyway. If you'd rather verify it than trust the screen, run Get-FileHash localgpu-setup.exe in PowerShell and compare against SHA256SUMS. It also needs administrator rights — it registers a system service.
It does not download any models. Unlike the Linux script, the Windows installer leaves your disk alone — until a model is present, the agent connects and serves nothing. Add one from the tray, no terminal needed: right-click the LocalGPU icon → Models → Pull a model…. The same menu lists what this machine has and marks what is actually being served.
02
Your hardware, your models
Requests are served by Ollama on your box. If the machine is offline, your workspace falls back to its configured provider automatically.
03
Bound to one workspace
The device enrolls with a short code and is bound server-side to the workspace that entered it. It never serves anyone else's traffic.
What the installer does
A few steps, then it's running.
On Linux
- Installs Ollama if it isn't already present, and pulls the starter models (llama3.1:8b and nomic-embed-text).
- Downloads the agent binary for your architecture and verifies its SHA-256 against a checksum baked into the script — a mismatch aborts the install.
- Installs it to ~/.local/bin/ollamaredirect-agent. No root required.
- Enrolls the device: it prints a short code. Open Magpy → Settings → Local GPU and enter that code to bind the machine to your workspace.
- Registers a systemd --user service so the agent starts on boot and survives logout.
Check on it later with systemctl --user status ollamaredirect-agent, or follow the logs with journalctl --user -u ollamaredirect-agent -f.
On Windows
- Checks for Ollama and points you at its download page if it's missing — it does not install it for you. It also pins the older CUDA backend automatically on pre-Turing cards, which would otherwise fall back to the CPU.
- Installs to C:\Program Files\LocalGPU and registers the LocalGPU service under its own low-privilege account — not SYSTEM. The service starts on boot and restarts itself if it ever fails.
- Starts the tray icon, which is the whole user interface: status, pause, logs, and a Models menu that lists what is installed, marks what is being served, and can add more.
- Enroll from the tray: right-click → Connect to uSaaS…. It opens Magpy and shows a short code; enter it under Settings → Local GPU.
- Pull a model — the installer downloads none. Right-click the tray icon → Models → Pull a model…, type a name from ollama.com/library, and it downloads and starts serving it. (ollama pull in a terminal works too; then use Models → Refresh list.)
The agent's log is at C:\ProgramData\OllamaRedirect\agent.log, or use Open logs in the tray menu.
Requirements
What you need.
A short list, and the installer handles the rest.
- Linux on x86-64 or arm64 (the installer targets Debian/Ubuntu), or Windows 10/11 on 64-bit hardware. Administrator rights are needed on Windows; on Linux the agent installs without root.
- An Nvidia GPU with the driver already installed (check with nvidia-smi). The installer does not install drivers.
- Outbound HTTPS/WSS to magpy.co. No port forwarding, no inbound rules.
- Owner or admin access on the Magpy workspace you're linking to.
- Enough VRAM for the model you choose — roughly 6 GB for a 7–8B model, 10 GB for 14B. A model that doesn't fit still runs, partly on the CPU, and is very slow.
macOS isn't supported yet. Tell us if you need it.
Manual download
Binaries and checksums.
If you'd rather not pipe a script to a shell, take the static binary directly. It's self-contained — no runtime dependencies. The Windows installer is the same file the button above links to.
| Platform | File |
|---|---|
| Linux · x86-64 | agent-linux-amd64 |
| Linux · arm64 | agent-linux-arm64 |
| Windows · 64-bit installer | localgpu-setup.exe |
| Checksums | SHA256SUMS |
Verify with sha256sum -c SHA256SUMS before running the binary — or, on Windows, Get-FileHash localgpu-setup.exe and compare. None of these files are code-signed yet.
Need something else
Need macOS, or help enrolling a device?
Email us — we can also walk you through setup live.
Contact support