How to Run Your Own AI Model at Home and Stop Paying for ChatGPT

What a self-hosted LLM setup gives you

Running a self-hosted LLM at home means installing a local AI model on your own hardware, adding a chat interface, and exposing it securely so you can use it from all your devices like a personal, private ChatGPT replacement without ongoing subscription fees.

If you are tired of paying every month for an AI assistant, or you want more privacy than a cloud service can offer, this approach is worth a look. A modest old desktop is enough; one user runs a 14B-parameter model comfortably on hardware that is no longer impressive, accepting slower replies for heavier models. Local AI models are not on par with top-tier cloud models for very hard reasoning, but they handle everyday coding help, writing, and planning tasks well enough to replace most casual queries. The real trick is making this free ChatGPT alternative feel as convenient as the original on your phone and laptop, which is where Tailscale remote access and tools like Ollama or Docker Model Runner come in.

What you need before you start

Before touching any settings, check that you have the right basics. First, you need a machine you can turn into a small home server, ideally an old desktop that can stay on when you need it. The reference setup uses a repurposed desktop that can comfortably run a 14B model and can handle something larger if you are patient with response times. You will also want at least one client device such as a laptop and a phone so you can feel the “ChatGPT everywhere” effect once remote access is working.

On the software side, you need either Ollama or Docker Model Runner as your self-hosted LLM engine, plus a chat front end like Open WebUI for a friendly interface. If you prefer Docker, you will need Docker Desktop 4.40 or newer on macOS, or 4.41 or newer on Windows, so that you can enable Docker Model Runner in the AI settings. Finally, install Tailscale on the server, your laptop, and your phone so they can all see each other over a private mesh VPN without exposing your network to the public internet.

Step-by-step: from bare box to free ChatGPT alternative

1. Prepare your home server and install a local AI engine: On your old desktop or home server, install either Ollama or Docker Desktop if you prefer Docker Model Runner. With Docker Desktop 4.40+ on macOS or 4.41+ on Windows, open Settings → AI and enable Docker Model Runner; you can also turn on host-side TCP access so other local apps can connect. This gives you a self-hosted LLM that can pull, run, manage, and serve local AI models. Ollama runs as a background service on the box and serves the local LLM in a similar way.
2. Add a ChatGPT-style web interface: Deploy Open WebUI and point it at your chosen backend so you get a chat experience instead of a bare terminal. With Ollama, Open WebUI talks to it directly and gives you conversation history, search, and file uploads so it feels a lot like ChatGPT’s web UI. With Docker Model Runner, set the OLLAMA_BASE_URL environment variable to the Model Runner endpoint; because it exposes an Ollama-compatible API, Open WebUI can treat it almost exactly like an Ollama server. Either way, you end up with a browser-based front end where models are still running locally on your machine.
3. Install Tailscale on every device: Next, install Tailscale on the server, your laptop, and your phone. Once you log into the same Tailscale account on all of them, they become reachable from one another over a private mesh VPN built on WireGuard, without any port forwarding or public exposure. This step is what turns your local AI model into a convenient tool you can reach from anywhere; after setup, one user reports they could reach their local LLM from anywhere in the same way they used ChatGPT, except conversations never left the house and no subscription fee was needed each month.
4. Give your AI server a friendly name and URL: In the Tailscale admin, enable MagicDNS and assign a readable name like “home-ai” to your server. Now you can open your browser on any device and go to a URL such as http://home-ai:3000 to reach the Open WebUI chat interface. At this point, functionality is there but browsers often complain that the site is insecure because it uses HTTP, which becomes annoying when you are checking your AI assistant throughout the day.
5. Secure the connection with Tailscale Serve and HTTPS: To silence browser warnings and add encryption, enable Tailscale Serve on your AI server. Tailscale Serve can issue a TLS certificate for your tailnet hostname, so your “home-ai” address loads over HTTPS without complex manual configuration. Once that is active, your phone and laptop treat the chat interface like any other secure site, while traffic still flows only inside your private Tailscale network. You now have a free ChatGPT alternative you can reach from everywhere you normally used cloud AI, but without paying for convenience or sending prompts to third-party servers.

The main trap here is to stop after getting a model running on the server. Without Tailscale remote access and a browser interface, it will feel like a toy you barely use, and you will fall back to your subscription out of habit. Adding friendly URLs, HTTPS via Tailscale Serve, and a web UI is what turns a local AI model into a tool you can rely on every day, not a weekend experiment.

Ollama vs Docker Model Runner for local AI

Both Ollama and Docker Model Runner give you a practical way to deploy a self-hosted LLM, but they fit slightly different styles. Ollama runs as a background service on your server and is designed to make “pull a model and chat” as straightforward as possible; one user has relied on it for a long time without needing many tweaks. It pairs naturally with Open WebUI, which then supplies a strong ChatGPT-like interface with history, Markdown rendering, model switching, and configurable generation settings.

Docker Model Runner, on the other hand, integrates models into the Docker world. Docker treats the model as its own OCI artifact while a dedicated inference engine such as llama.cpp or vLLM runs it locally, with the model cached on disk after the first download. For most users, the default llama.cpp backend matters most because it runs quantized GGUF models across Apple Silicon, Windows, and Linux, and it supports automatic Metal acceleration on capable Macs. You can talk to Model Runner over both OpenAI-compatible and Ollama-compatible APIs, so many tools built for Ollama can connect by changing the base URL, including coding assistants like Cline, Continue, Aider, Cursor, and OpenCode.

Privacy, costs, and whether this swap is worth it

Once your phone can reach your home AI as easily as it reaches ChatGPT, the case for a subscription shrinks. One user found that when their local LLM had the same universal access as cloud AI, “there wasn't much reason left to keep paying.” Local AI models that run well on modest, outdated hardware have not caught up to top cloud models for the hardest problems, but most daily questions are low-stakes and only need basic answers, making ongoing fees hard to justify.

The privacy gains are often the bigger win. When the LLM lives entirely on your own server, none of your conversations leave your house and nothing you type touches a third-party server. That makes it easier to feed your assistant drafts, logs, or personal notes without worrying about where they end up. The hardware requirements are modest, since even a no-longer-impressive desktop can run sizable models as a local AI model, and once the system is in place, you are no longer trading convenience for a monthly fee. The thing to watch is your expectations: treat this setup as a dependable free ChatGPT alternative for everyday work, not as a perfect clone of the newest cloud model.