Self-hosted local LLM stack: Ollama (model runner) + Open WebUI (ChatGPT-style frontend). The agent pulls the model you select on first install and serves the WebUI on the bound domain.
Notes
On a Pi 5 with 8 GB RAM, 4B-parameter quantised models like qwen3:4b or llama3.2:3b produce a few tokens per second on CPU and are usable for personal tinkering. Anything larger will swap.
Ollama always exposes its OpenAI-compatible API on obacht-<instance>-ollama:11434/v1 inside the obacht-edge network, so other instances on the same device (n8n, custom apps) can consume it via a service_reference config field of interface openai_compatible.
Toggle expose_ollama to true if you also want to bind a separate domain to the raw Ollama API for use from outside the device — otherwise only the WebUI is publicly reachable.
Heavy: needs ~3 GB RAM idle, several GB of disk per model. The models volume is preserved across reinstalls.
Open-source contribution
A fixed amount per install goes to Open WebUI, forwarded via GitHub Sponsors.
Ollama in this bundle cannot take money. The contribution goes in full to Open WebUI.
Last checked 2026-07-29
01
Connect a device: a Raspberry Pi, a Linux machine, or an Apple silicon Mac.
02
Click install in the dashboard. The template is signed; your device verifies the signature before anything runs.
03
Optionally attach your own domain. HTTPS certificate included, no port forwarding.
Your data stays on your device. obacht handles the tunnel, SSL, and domain.
Sign up for the closed beta