I’m changing how I use AI (Open WebUI + LiteLLM)

A single self-hosted dashboard can unify access to many major AI models—cloud-hosted chat systems like GPT and Claude, plus local models—while letting an owner control who can use what, set budgets, and monitor or restrict prompts. The approach centers on Open WebUI (an open-source, self-hosted web interface) paired with LiteLLM (used as an API-compatible proxy/gateway), turning scattered subscriptions and separate logins into one “AI hub” with admin controls.

Open WebUI is positioned as the front end: it runs on a machine you control (either a cloud VPS or on-prem hardware such as a laptop, NAS, or Raspberry Pi) and can connect to “whatever LLM” you choose. In the cloud setup, the workflow uses a VPS from a hosting provider, then installs Open WebUI alongside Llama (a local model) on Ubuntu 24.04. Once deployed, the admin creates an initial admin account, then can chat with a default local model (Llama 3.1 2B in the walkthrough) through the familiar Open WebUI interface—typically slower than the largest cloud models, but dependent on your hardware.

The key unlock is model access beyond local Llama. For cloud models, the transcript explains that Open WebUI can connect to providers via APIs, which is often cheaper and more flexible than paying for multiple full “normie” chat plans. After creating an OpenAI API key, the key is pasted into Open WebUI’s admin settings under connections, enabling access to a range of OpenAI models—including newer ones such as “4.5” mentioned in the walkthrough—inside the same interface.

But cost control is the major caveat. Billing for text LLM usage is described as token-based: tokens are roughly word fragments and punctuation, and pricing varies widely by model. The transcript gives example rates (e.g., $1.10 per 1M tokens for “o3-mini,” $15 per 1M tokens for “o1 reasoning,” and up to $75 per 1M tokens for “4.5,” with input costs called out). Conversation length and “context” increase token usage because prior messages are repeatedly sent, so heavy use can quickly exceed expectations. Caching is mentioned as a potential cost reducer, but not a guarantee.

To solve Open WebUI’s limited built-in connection types (only OpenAI and a local option are shown), LiteLLM is introduced as the proxy layer. Open WebUI connects to LiteLLM using an OpenAI-compatible API, while LiteLLM fans out to many other providers (Claude, Gemini, Grok, DeepSeek, and more). LiteLLM is deployed via Docker, configured with master and encryption keys, then set up with provider API keys. Within LiteLLM, “virtual keys” can restrict which models a given user can access and can enforce monthly budgets.

Finally, Open WebUI’s user and group controls are used to apply guardrails for family and employees. The walkthrough creates a “kids” group, assigns users, limits accessible models (e.g., only Claude 3.7), and adds a system prompt instructing the assistant to help with schoolwork without doing it for them. Admins can also review chat logs for monitoring, and the transcript notes that chat history visibility can be turned off for kids. A separate follow-up is teased for setting up a friendly DNS name instead of exposing an IP address.