openHAB MCP Server + Ollama + Open WebUI — Complete Setup Guide Hardware Recommendations To run a capable LLM model alongside openHAB MCP and Open WebUI, the following hardware is recommended: Minimum (CPU-only, small models only) CPU: Modern x86_64 with AVX2 support, 8+ cores on a single NUMA no…

Tool calling with a LLM is completely a demonic hack. (demonic as “how the hell can something like this really work at all??”). But a hack that can do incredible thing nonetheless. Many small models mess this up, and do not put the correct JSON markup to do ‘proper’ tool calling. Then JSON garbage …

[image] dalgwen: But it’s surprising, because the qwen model you use (32b) should be pretty good at this. Yes, it’s not like I’m trying to use a 3b model.. Just as a test, I switched to llama3.1:8b-instruct-q8_0, and then I actually get a response, not just JSONs.. Based on the tool call r…

To get a more capable model that isn’t Qwen, I’ve downloaded mistral-small. It, too, can talk to the MCP, but I’m wondering if it has something of a “God complex”: [image] I really hope I don’t accidentally make it turn off the sun!

The usual “AI slop” is prevalent. I asked it a stupid question, but it was one of the “suggested follow-ups”. Is there a way to automatically restart or reconnect these offline devices? The answer is that there’s nothing it can do to reconnect these, most of them don’t even exist (they are just Th…

Hey! Thanks for the guide, even though I didn’t use it to its full extent. I launched Ollama with Open WebUI via a Docker container, which is less of a hassle. The connection with openHAB seems to have worked as well. To play around with the hardware a bit, I tried getting the LLM to run on my R…

[image] Hoerli: I think the built-in chat feature in OH 5.2 can be very useful, but I don’t want to connect to Google. I want everything to run locally on my end. Give the ChatGPT binding version on the marketplace a try, it supports LLM-based HLIs and allows to connect to any OpenAI-compati…

[image] Hoerli: I tested the following: qwen2.5:0.5b → Makes up too many things and wants to do things it’s not supposed to do qwen2.5:7b → CPU constantly at 100%, nothing happens qwen3.5-0.8b → CPU constantly at 100%, nothing happens qwen3.5:2b → CPU constantly at 100%, nothing happens mis…

TBH I’d just go with Mistral AI or Hetzner inference (EU-based), or any other EU-based Hoster providing larger open weights models. I don’t see a situation in the near future where you could run a large enough model at home without crazy hardware investments … Eventually your university or company…

Thanks for the comments @Nadahar and @florian-h05 . I just wanted to test whether the smaller models are sufficient for simple commands. I didn’t want to create scripts or adjust configurations just (yet). I’d rather do that myself (for now). For simple tasks, like “How warm is it in room XY?” o…

[image] Hoerli: My smart home will never be connected to public clouds. That’s my standard, and I stick to it. I agree with you 100%, which is why I’ve been experimenting with local models. I have a GPU with 11GB VRAM, so I have a little bit of room to maneuver, but I haven’t really found an…

openHAB MCP Server + Ollama + Open WebUI — Complete Setup Guide

Tutorials & Examples Solutions

florian-h05 (Florian H.) August 16, 2026, 6:43pm 19

Give the ChatGPT binding version on the marketplace a try, it supports LLM-based HLIs and allows to connect to any OpenAI-compatible API.