SaaS chat subscriptions add up—ChatGPT Plus, Claude Pro, and team seats can exceed $20–30 per user per month while your prompts and uploaded files live on someone else's servers. Open WebUI (MIT-licensed, 110,000+ GitHub stars as of June 2026) gives you a polished, ChatGPT-like web interface that runs on your Mac: connect Ollama for free local models, or route to OpenAI-compatible APIs (OpenRouter, DeepSeek, Groq) while keeping the UI and document library under your control.
For beginners who want privacy + RAG without wiring LangChain by hand, Open WebUI ships built-in Retrieval Augmented Generation: upload PDFs and Markdown to a document library, then reference them in chat with the # command before your question. This guide walks through a one-command Docker install on macOS, first model pull, and a minimal knowledge-base workflow—no NodeMac pricing pitch, just upstream paths from the official docs.
Wie Open WebUI und Ollama zusammenpassen
Open WebUI is the browser-facing layer; Ollama (or a remote OpenAI-compatible endpoint) is the inference engine. Persistent state lives in a Docker volume mounted at /app/backend/data.
Browser → http://localhost:3000 (or :8080 with --network=host)
→ Open WebUI container (FastAPI + Svelte UI)
→ Ollama at host.docker.internal:11434 OR OPENAI_API_BASE_URL
→ Vector DB (default embedded) + uploaded docs in data volume
→ RAG: user types "#" + selects document collection before prompt| Component | Default location / config | Role |
|---|---|---|
| Web UI | Port 3000 mapped to container 8080 | Chat, admin, model picker |
| Ollama | 11434 on host or bundled in :ollama image | Pull/run llama3, qwen2.5, etc. |
| Document library | Admin → Documents / Workspace | PDF, TXT, MD for RAG |
| RAG trigger | # in chat input | Attach knowledge to a single turn |
| Data volume | -v open-webui:/app/backend/data | Required—stores DB + uploads |
Quotable: Upstream warns that omitting the open-webui named volume wipes users, chats, and uploaded documents on container recreate—always mount -v open-webui:/app/backend/data.
Quelle: Open WebUI README and getting-started docs.
ChatGPT Plus vs self-hosted Open WebUI
| Factor | ChatGPT Plus (~$20/mo) | Open WebUI + Ollama (local) |
|---|---|---|
| Data residency | OpenAI servers | Your Mac / your Docker volume |
| Document RAG | GPT store / limited uploads | Full doc library + # per chat |
| Model choice | OpenAI models only | Any Ollama model + API backends |
| Offline use | No | Yes, with local models pulled |
| Setup time | 0 minutes | ~15–30 minutes first install |
| Laufende Kosten | Abo | Strom + optionale API-Keys |
If you only need occasional GPT-4o quality, do point Open WebUI at a cheap OpenAI-compatible API and skip heavy local models. If you want zero cloud inference, do use the bundled :ollama image and models like llama3.2:3b on a 16 GB Mac—see Apple Mac mini specifications for baseline RAM.
Schritt-für-Schritt-macOS-Runbook
- Install Docker Desktop — Open WebUI's recommended path for Mac beginners. Enable WSL-like VM resources: allocate 8 GB+ RAM to Docker if you run 7B+ models.
- Choose install flavor — Three common patterns:
# A) Open WebUI only — Ollama already running on Mac (brew install ollama && ollama serve)
docker run -d -p 3000:8080 --add-host=host.docker.internal:host-gateway \
-v open-webui:/app/backend/data --name open-webui --restart always \
ghcr.io/open-webui/open-webui:main
# B) Bundled Open WebUI + Ollama (simplest one-container start)
docker run -d -p 3000:8080 -v ollama:/root/.ollama -v open-webui:/app/backend/data \
--name open-webui --restart always ghcr.io/open-webui/open-webui:ollama
# C) Connection issues? Use host networking (note port becomes 8080)
docker run -d --network=host -v open-webui:/app/backend/data \
-e OLLAMA_BASE_URL=http://127.0.0.1:11434 --name open-webui --restart always \
ghcr.io/open-webui/open-webui:main- Open the UI — Visit http://localhost:3000 (or http://localhost:8080 with
--network=host). Create the first admin account—registration closes after the first user on private installs. - Pull a starter model — In another terminal (if using separate Ollama):
ollama pull llama3.2:3b
# or for Chinese/English mix: ollama pull qwen2.5:7bBundled :ollama image: use Admin → Settings → Models or docker exec -it open-webui ollama pull llama3.2:3b.
- Select model in chat — Top bar → pick
llama3.2:3b. Send a test prompt to confirm Ollama connectivity. - Build document library (RAG) — Admin → Documents (or Workspace files): upload PDFs,
.md,.txt. Wait for embedding to finish (status in UI). - Query with
#— In a new chat, type#, select your collection or file, then ask: "Summarize refund policy section 3." Open WebUI injects retrieved chunks into the prompt. - Optional: OpenAI-compatible API — Settings → Connections: set base URL (e.g. DeepSeek, OpenRouter) + API key for cloud models while keeping the same UI and doc library.
Alternative (pip): pip install open-webui && open-webui serve runs on http://localhost:8080 without Docker—upstream requires Python 3.11.
Fehlerbehebung
"Open WebUI: Server Connection Error" (Ollama unreachable)
Symptom: Models list empty; error mentions 127.0.0.1:11434 inside container.
Fix: Docker on Mac cannot reach host Ollama via bare localhost. Use --add-host=host.docker.internal:host-gateway (pattern A) or --network=host (pattern C). Verify: curl http://127.0.0.1:11434/api/tags on the host shows models.
Wrong port after install
Symptom: Browser cannot connect.
Fix: Default mapped install uses host port 3000. --network=host switches to 8080. Pip install also uses 8080. Match your URL to the command you ran.
Documents uploaded but # returns nothing
Symptom: RAG attach succeeds but answers ignore file content.
Fix: Confirm embedding completed (no error badge on document). Retry with a smaller PDF. Check Admin → Settings → Documents / RAG: ensure the collection is included in chat and the model supports context length.
Container recreate lost all chats
Symptom: Fresh install after docker rm.
Fix: You omitted or deleted the open-webui volume. Always use -v open-webui:/app/backend/data. Inspect with docker volume inspect open-webui.
Weiterführende Artikel
Kombinieren Sie die Open-WebUI-Dokumentenbibliothek mit OpenHuman Mac mini M4 OAuth-Speicher oder Understand-Anything-Wissensgraphen auf demselben always-on Mac mini.
Beim Modell-Pinning in Open WebUI hilft unser
GPT-5.6 iris-alpha Codex-Leak-Tracker
—gpt-5.5 pinnen bis die Systemkarte erscheint.
Apple Intelligence auf dem Mac verfolgen? Unser WWDC-2026-All-Systems-Glow-Siri-Leak-Decoder —Evidenzmatrix, 8 Schritte, 5 FAQ.
FAQ
Ist Open WebUI dasselbe wie OpenAIs ChatGPT?
Nein. Open WebUI ist eine Open-Source-Self-Host-UI für Ollama, OpenAI API oder andere kompatible Backends. Sie steuern Modelle, Nutzer und Dokumentenspeicher selbst.
Brauche ich eine GPU auf Mac mini M4?
Nicht für kleine 3b–7b quantisierte Modelle auf Apple Silicon—Ollama nutzt Unified Memory. 16 GB reichen für llama3.2:3b; 24 GB bei 7b + Embedding parallel. :cuda-NVIDIA-Images sind für Linux/NVIDIA, nicht Mac Docker.
Komplett offline nutzbar?
Ja nach ollama pull ohne Cloud-APIs. HF_HUB_OFFLINE=1 blockiert HuggingFace laut upstream README.
Wie unterscheidet sich RAG vom ChatGPT-Upload?
Open WebUI speichert in Ihrer Dokumentenbibliothek mit Vektor-Backends (ChromaDB, PGVector). Der #-Befehl hängt Sammlungen pro Nachricht an—ideal für Multi-Projekt-Wissensbasen.
Docker zu schwer auf dem Laptop?
pip install auf einem always-on Mac mini oder Docker auf einem SSH-Host. PWA im LAN unter http://your-mac-ip:3000—kein Port-Forward ohne Auth-Härtung.