Open WebUI: A ChatGPT-Style Interface for Your Local Models
Note: Some future links on this site may be affiliate links. We never claim hands-on testing; guides are researched from the public sources listed in each article.
How we wrote this: A summary of Open WebUI's official documentation and GitHub README (linked below). We have not benchmarked it. Version numbers and image sizes are as the docs state them in October 2026 and will change.
Ollama and llama.cpp run models, but their built-in interfaces are basic. Open WebUI is a self-hosted web app that adds a polished chat interface on top: conversation history, multiple users, document chat and more. It runs on your own machine.
What it is
The Open WebUI docs describe it as a self-hosted AI platform "built to run entirely offline". It works with Ollama and with "OpenAI-compatible APIs". That covers a local Ollama install, a llama.cpp server, or a cloud provider if you choose one. Per the quick start, it runs on macOS, Linux and Windows, and Docker is the "officially supported and recommended" install method.
The Docker quick start, explained
The quick start's command maps ports, storage and a secret key. Here's what each part does, per the docs:
| Part of the command | What it does |
|---|---|
-p 3000:8080 |
The interface opens at http://localhost:3000 on your computer. |
-v open-webui:/app/backend/data |
Stores your chats, users and settings in a Docker volume that survives updates. The docs say to "never run without it". |
--add-host=host.docker.internal:host-gateway |
Lets the container reach Ollama running on the same computer. |
--restart always |
Starts it again after a reboot. |
-e WEBUI_SECRET_KEY=... |
A fixed key. Without one, recreating the container logs everyone out. The docs suggest generating it with openssl rand -hex 32. |
There's also a pip install (the README asks for Python 3.12) and a desktop app. For long-term use, the docs recommend Docker or Python.
Which image: main or slim
Per the quick start:
:mainis the standard image. It bundles speech-to-text and embedding models, so document chat and voice work inside the container.:slimis about 197 MB to download instead of 1.62 GB. The docs measured roughly 300 MiB of RAM at idle for it. The chat experience is the same, but document search and voice need separate services.:ollamabundles Ollama inside the container for an all-in-one setup.:cudaadds Nvidia GPU support.
If you mainly want chat with a local Ollama, slim keeps more memory free for the model itself. That matters because the model, not the interface, is usually what strains a laptop. Private AI Computer's memory calculator estimates model memory.
Connecting to Ollama
With Ollama running on the same computer, the host.docker.internal setting lets Open WebUI find it. The README's troubleshooting section says connection problems usually mean the container can't reach Ollama at 127.0.0.1:11434. Its suggested fix is the --network=host flag, which changes the address to http://localhost:8080. Our Ollama API guide explains that port and the server behind it.
Keeping it private
- Offline mode. The README says setting
HF_HUB_OFFLINE=1stops attempts to download models from the internet. - Accounts and roles. The README describes role-based access control, so in a shared office you can limit what each user can do.
- Network exposure. Port 3000 on your computer may be reachable from your local network. Don't forward it from your router to the internet.
- Cloud connections are optional. If you add a cloud API key, chats sent to that model leave your machine. Stick to local models for anything sensitive.
For the bigger picture on keeping AI work on your own hardware, see chatting with your documents locally and Private AI Computer's what private AI means.
Licensing note
The README says the current codebase includes components under the Open WebUI License, which has an extra requirement to keep the "Open WebUI" branding. Read the LICENSE file before rebranding it for clients. Our model licenses guide covers the separate question of model licenses.