Cloud AI subscriptions keep getting more expensive, and every prompt you type is data you no longer control. Self-hosting your own AI assistant on a virtual private server flips that arrangement around: you rent a small slice of a data center, run an open-weight model on it, and chat with it from your phone or laptop with nothing leaving your machine.

It is cheaper than it sounds. A basic VPS with enough memory for a 7 to 8 billion parameter model costs roughly five to ten dollars a month, and free open-source chat interfaces make the raw model feel like a familiar chatbot. This guide walks through the full setup step by step, assuming nothing beyond a willingness to paste commands into a terminal.

Any mainstream VPS provider works: Hetzner, DigitalOcean, Linode, and Hostinger all offer suitable plans. Aim for at least 4 GB of RAM for a 7B model, though 8 GB gives the model more breathing room. Choose Ubuntu 22.04 or 24.04 as the operating system, since nearly all AI tooling is documented against it. Set aside about an hour of uninterrupted time.

Log in over SSH with the credentials your provider sent you, then update the package lists and upgrade the system software. This takes a few minutes on a fresh machine and closes known security holes before you install anything else.

If you landed as the root user, create a regular user account now and enable the built-in firewall with only SSH, HTTP, and HTTPS ports open. Securing the box before installing software is a habit that pays for itself many times over.

Ollama is the simplest way to download and run open-weight models on a server. Its one-line installer detects your system, sets up the service, and starts it in the background. Verify the install with the version command; if a version number appears, the engine is alive and listening on the local interface only, which keeps the model API off the public internet for now.

Pull a mid-size instruction-tuned model such as Llama 3.1 8B or Gemma 2 9B. The download is several gigabytes, so be patient on slower connections. When it finishes, run the model once from the terminal and ask it a simple question to confirm everything works.

On a 4 GB VPS, lighter models like Phi-3 or Gemma 2B fit more comfortably, at the cost of noticeably weaker answers on hard questions. Start small - upgrading a VPS later is usually a one-click operation.

A terminal prompt is fine for testing, but a web interface is what makes the assistant usable every day. Open WebUI is the most popular choice: it runs in a Docker container, talks to Ollama, and gives you chat history, multiple conversations, and a clean layout that works on mobile browsers.

Install Docker, then launch the Open WebUI container with a persistent volume so your chats survive restarts, mapped to a high port like 3000. Open your server's IP address in a browser and you should be chatting with your model within minutes.

Never leave the chat interface open to the internet without authentication. Place Nginx or Caddy in front of it as a reverse proxy, get a free TLS certificate from Let's Encrypt, and enable the built-in login controls so only your account can sign in.

Then tighten the firewall so the chat port is no longer reachable directly, leaving only ports 80 and 443 open. If you want to be extra careful, restrict access to your home IP address or require a VPN connection. Your prompts now travel over encrypted connections to a server only you control.

A 4 GB VPS runs about five dollars a month; 8 GB is closer to ten or twelve. There are no per-token charges and no usage tiers - heavy months cost exactly the same as light ones. Compared with a twenty-dollar monthly chatbot subscription, self-hosting breaks even quickly and gets cheaper the more you use it.

The trade-off is maintenance: you handle updates, backups, and the occasional troubleshooting session. For many people that is a fair price for privacy, because your conversations, documents, and code never sit on someone else's servers.