I got tired of the ChatGPT export. You know the one, the "download your history" button that exists so you can technically leave if you want to. I used it four times in a year, and every time it just confirmed that my prompts, my threads, my half-finished code sat on someone else's servers with my name on them.
So when I started messing with local models, the first thing I actually wanted wasn't raw inference. Raw inference is fun for an afternoon, then you realize you miss the chat UI. The history. The threads. The context.
That's where Open WebUI came in. It's a self-hosted web interface that looks and feels like the chatbot you're used to. It runs in Docker, connects to whatever model backend you've got running, and keeps all your chats in a local database. Not a vendor's cloud. A database on your machine.
Honestly, setup is a single container. The bare version of it is:
docker run -p 3000:8080 --name open-webui ghcr.io/openwebui/open-webui
Point it at your model backend, sign in once, and suddenly you've got a chatbot that doesn't phone home. My prompts stay in my house. My threads stay in my house. If my internet blips at 9pm, mine still answers, because nothing needs to leave the machine. That's the whole pitch in one sentence.
What actually sold me is the flow of it. I do a fair bit of support work and admin, so I ask a lot of small questions. Sending each of those to a hosted chat means every one flows through a third party and lands in someone's logs. With my own web UI pointed at a local quantized model, the cost of a trivial question is basically zero, so I ask more of them. That's the part nobody tells you about self-hosting: you use the AI more, not less, because asking never feels like a leak.
Now the honest part. Local models have limits, and I won't pretend otherwise. A machine you own runs smaller models than the frontier ones do. But a well-quantized 7-9B model handles my typical jobs fine, drafting and summarizing and gluing things together. It's only when I need deep reasoning that I feel the gap. The trick is treating it like any other tool and matching the size to the job.
If you want a home lab without assembling the pieces yourself, I put together the Self-Hosted AI Home Lab as a bundle, pay what you want from nine dollars. That's LM Studio, Open WebUI, Docker, and a set of quantized models, all self-hosted, no cloud, ready to tinker with. https://symshah.gumroad.com/l/selfhosted-ai-homelab
I'll take a chat that never leaves my house over renting one that does. The history button can wait forever.
Top comments (0)