Michael Limberger

The same window, on hardware you own

This talk is a practical OpenWebUI setup. You get a familiar chat window talking to Ollama on your network, without sending the conversation to a public cloud chat.

Tonight we build that experience on your own machine. OpenWebUI is an open source web interface that looks and feels like ChatGPT. It connects to Ollama, which runs the models locally. Private, under your control, and powerful enough for real work once the models are on disk.

Why bother

Three reasons show up again and again when people stay local: privacy, freedom, and cost. Privacy comes first for most of us. Conversations, documents, and questions stay on the machine. No public terms of service harvest, and no mystery about where the text went.

Freedom is next. You are not stuck with someone else's content filters, and you do not get a canned "I can't help with that" on a reasonable question. You pick the model. You set the rules.

Cost is the quiet third reason. After setup there is no monthly subscription and no per-token bill. The only ongoing cost is the electricity your Mac uses.

OpenWebUI itself is a web app on your computer. You open it at localhost:3000. You get chat, history, file upload, and a model selector at the top. Under the hood it talks to Ollama through an API. You type, OpenWebUI sends the request, Ollama runs the model on your hardware, and the response streams back.

+------------------------------------------+
|  Your Mac                                |
|                                          |
|  Browser -----> OpenWebUI -----> Ollama  |
|  (localhost)    (Docker)         (local) |
+------------------------------------------+

Everything stays on the machine. Once it is installed, you do not need the internet to chat.

What we cover

We install and run OpenWebUI with Docker, connect it to Ollama, and pull models. We navigate the chat UI, create custom agents in Workspace, write system directives, and tweak temperature and top-p. We also try vision models on images and walk through troubleshooting. Along the way we look at several models, what hardware they need, and when you would pick one over another.

Hardware target: an Apple Silicon Mac (M1, M2, M3, or M4). 16GB unified memory is the minimum. 32GB is more comfortable. Plan on about 30GB free disk. Software: Colima plus the Docker CLI, and Ollama. If you were at the Claude Code session two weeks ago, you probably already have Ollama. You are ahead.

This is a casual walkthrough, not a certification. We set things up, poke around, and have some fun. If something breaks on your machine, we debug together or you can ask after. The goal is zero to a working local ChatGPT-style setup. Next: the hardware check.