Private ChatGPT for a small business (Open WebUI + LiteLLM)

Components

• Open WebUI: Chat interface, user accounts, document upload
  Official hardware: Not published (measured by us: 909 MiB RAM while working) · License: Open WebUI License (BSD-3-Clause style with branding clause) (Custom / conditional)
• LiteLLM: Model gateway: API keys, model access control, budgets, spend logs
  Official hardware: minimum 1 vCPU · 4 GB RAM · License: MIT (core) + LiteLLM Enterprise License (enterprise directory) (Open core)
• Qdrantoptional: Vector database for document search (instead of the built-in default)
  Official hardware: Not published (measured by us: 566 MiB RAM while working) · License: Apache-2.0 (Permissive)

Combined hardware and monthly cost

How we got this: we added up the official minimum vCPU and RAM each project publishes (linked on its app page) and took the cheapest Hetzner Cloud Regular Performance (shared AMD vCPU) plan that meets the total. Prices read from www.hetzner.com/cloud/regular-performance on 2026-10-22. "From" monthly price cap shown by Hetzner, excluding VAT. Prices vary by location. Covers the server only; it excludes LLM API usage, backups, extra storage, GPUs and your time. Running everything on one server is our simplifying assumption. Measured figures, where we have them, are on each app page.

Cost vs ChatGPT / Claude / Microsoft 365 Copilot

Deployment outline

1. Deploy the LiteLLM proxy with a config file listing the models you want to offer, and a master key. source
2. Give LiteLLM a database. Without one, virtual keys do not work and budgets are not enforced. source
3. In the LiteLLM UI (port 4000, /ui) create a Team and a Virtual Key, and limit the key to the models Open WebUI may use. source
4. Deploy Open WebUI (official Docker command on its app page). source
5. In Open WebUI go to Settings → Connections and add a connection with the LiteLLM base URL and the virtual key. source
6. For per-user spend tracking, set ENABLE_FORWARD_USER_INFO_HEADERS=True in Open WebUI and map the X-OpenWebUI-User-Id or -Email header in LiteLLM's general_settings.user_header_mappings. source
7. Optional: to use Qdrant, set VECTOR_DB=qdrant and QDRANT_URI (plus QDRANT_API_KEY) in Open WebUI's environment. source

This is a summary of the official docs, not a full tutorial. Each app page has the exact official install command.

Gotchas

• Open WebUI's default vector store (ChromaDB on SQLite) is not safe for multi-worker or multi-replica deployments. Switch to a server vector DB such as Qdrant or pgvector before you scale out. source
• Users can change their own Open WebUI usernames and admins can change emails, so pick the user-ID header if spend reports must be tamper-proof. source
• Open WebUI: Branding may only be changed for deployments of 50 or fewer end users in a rolling 30-day window, unless you have permission or an enterprise license. source
• Open WebUI: If you only use hosted APIs, the :slim image is about 197 MB instead of 1.62 GB. It cannot be combined with the CUDA or bundled-Ollama variants. source
• LiteLLM: Without a database, budgets are not enforced. max_budget is silently ignored apart from a one-time startup warning, and virtual keys do not work. source
• LiteLLM: Memory is not a usable autoscaling signal for the proxy; scale on CPU instead. source
• Qdrant: The quick-start command runs without authentication and listens on all interfaces. Secure the instance (API key, network) before real use. source

What this does not replace

• None of these apps includes an AI model. You connect a model API (billed by its provider) or run open models yourself, for example with Ollama.
• Answer quality depends on the model you connect, not on the chat app.
• You are responsible for the server, updates, backups and user management.
• None of these works inside Word, Excel, Outlook or Teams the way Copilot does; they run as separate web apps.
• You connect your own model (API or self-hosted).

Other stacks

• Private document Q&A over your company's knowledge (Onyx)
• Self-hosted AI workflow automation (a Zapier replacement with n8n)
• Self-hosted meeting transcription (Speaches + Open WebUI)
• Self-hosted coding assistant for your dev team (Tabby)
• Self-hosted customer support AI chatbot (Dify)
• Fully offline private AI chat (Ollama + Open WebUI)

Not sure which app fits? Try the 5-question wizard.