Self-hosted AI stacks
A business rarely needs one app. It needs a working combination. Each stack lists the components from our catalog, what they need together on one server, and what that server costs at Hetzner Cloud.
Private ChatGPT for a 20-person company
Open WebUI + LiteLLM + Qdrant (optional) A self-hosted ChatGPT-style portal for a 10–50 person team. Open WebUI as the chat interface, LiteLLM as the gateway that holds API keys, sets budgets and logs spend per user. Optional Qdrant vector database. Server from €19.99/mo (CPX22), floorPrivate document Q&A over company knowledge
Onyx (formerly Danswer) + LiteLLM (optional) + Ollama (optional) Let staff ask questions across company documents and workplace apps with cited answers. Onyx indexes your sources; LiteLLM or Ollama supply the model. Server from €69.99/mo (CPX42)AI workflow automation (Zapier replacement)
n8n + Qdrant (optional) + Ollama (optional) Run Zapier-style automations with AI steps on your own server. n8n for workflows and AI agent nodes; optional Qdrant for document search and Ollama for local models. Server size not estimated (no official figures)Meeting transcription and summaries
Speaches + Open WebUI Transcribe recordings in-house with Speaches (faster-whisper, OpenAI-compatible API) and use Open WebUI to transcribe voice input and summarise with the model of your choice. Server size not estimated (no official figures)Local coding assistant (GitHub Copilot alternative)
Tabby + Ollama (optional) Code completion and chat served from your own GPU server with Tabby, optionally using models served by Ollama. Code never leaves your network. GPU server (not estimated)Customer support AI chatbot
Dify + LiteLLM (optional) Build a knowledge-base chatbot on Dify, embed it on your website, and route model calls through LiteLLM for spend control. Server from €19.99/mo (CPX22)Fully offline private AI (no cloud API)
Ollama + Open WebUI Run open-weight models on your own hardware with Ollama and give staff a ChatGPT-style interface with Open WebUI. No data goes to a model provider. Server size not estimated (no official figures)Server sizes add up each component's official minimum CPU/RAM. "Floor" means at least one component publishes no figure, so the real need is higher. Prices exclude VAT and model API usage. Compare against your SaaS bill with the cost calculator.