Create a desktop AI powerhouse that works without any internet connection — complete hardware assembly, operating system, models, and interfaces.
An offline AI workstation is the ultimate expression of self-hosted AI privacy. Unlike a home server that's connected to your network (and therefore, indirectly, to the internet), an offline workstation operates in complete isolation. It never phones home because it has no home to phone.
This guide builds a high-performance desktop machine running AI without internet connection. Every model, every inference engine, every interface — all installed from downloaded files, all running without a single packet leaving the machine. This is the setup that journalists, researchers, legal professionals, and privacy absolutists demand.
An offline workstation has slightly different requirements than a server. You need a display, input devices, and enough local storage to hold your entire model library and operating system — because you can't stream or download anything later.
Assembly follows standard PC building practices, with a few offline-specific considerations:
Since the final machine will be offline, you need to prepare all software on a separate internet-connected computer:
ollama.com/install.sh. Save it.~/.ollama/models/ — copy the entire directory to your USB drive.docker pull ghcr.io/open-webui/open-webui:main) and export it (docker save).Label this USB drive clearly — it's your offline software repository. Keep it updated as new models and tools are released.
sudo bash -c "echo 'blacklist nouveau' >> /etc/modprobe.d/blacklist-nvidia-nouveau.conf" and sudo update-initramfs -usudo systemctl stop gdmsudo bash NVIDIA-Linux-*.run --no-opengl-files (the --no-opengl-files flag prevents conflicts with the system's Mesa libraries)sudo rebootAfter reboot, verify with nvidia-smi. You should see your GPU with driver version and CUDA version displayed.
sudo bash install.shsudo systemctl stop ollama~/.ollama/models/sudo systemctl start ollamaollama list should show all your imported models.ollama run llama3.1:8b "Hello, are you running offline?"At this point, you have a working command-line AI assistant with zero internet connectivity. This is the core of your offline AI device — everything else builds on this foundation.
Install Open WebUI from your USB drive's Docker image:
sudo dpkg -i docker-*.debdocker load < open-webui.tardocker run -d -p 3000:8080 --add-host=host.docker.internal:host-gateway -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:mainOpen Firefox and navigate to http://localhost:3000. Create your admin account. You now have a ChatGPT-style interface backed entirely by your local hardware, with no internet connection whatsoever.
For a more appliance-like experience, a ClawBox private AI device provides this exact setup in a pre-configured box — ideal if you want the privacy benefits of offline AI without the multi-hour build process. But if you've followed this guide, you've built something more powerful and more customizable for less money.
Your offline workstation isn't just for chatting — it's a complete development and productivity environment. Install from your USB drive:
sudo dpkg -i code-*.deb — install the Continue extension (also downloaded offline) for AI-assisted coding entirely on-device.Having built the workstation without internet, now ensure it stays that way:
sudo ip link set eth0 downsudo ufw default deny incoming && sudo ufw default deny outgoing && sudo ufw enableYour workstation is now a true air-gapped system. It runs AI completely isolated from any network — the gold standard for running AI offline.
An offline machine requires a different maintenance workflow:
For a networked alternative that still prioritizes privacy, see our home AI server setup guide. For help picking the right components, check the hardware selection guide. And if you want to understand what software to run on this beast, the software and models guide has you covered.
Building an offline AI workstation is the ultimate project for anyone serious about private AI hardware. It's more work than buying a pre-built box or cloud subscription, but the result is something no company can take away from you: an AI that is truly, completely, permanently yours.
Always have a software USB: Keep a dedicated USB drive with drivers, models, and tools. Update it during your maintenance windows.
Test before air-gapping: Verify everything works while still connected. Troubleshooting is much harder offline.
Document everything: Write down your configuration in a local text file. You can't Google solutions from an air-gapped machine.