Every prompt you send to a hosted AI service is a document you handed to someone else, along with whatever you pasted into it. Self-hosting removes that: Ollama runs the model on your server, Open WebUI gives you the familiar chat interface, and the conversation never leaves infrastructure you control.
Two pieces that fit together
- Ollama runs the models: Llama, Mistral, Phi, DeepSeek and more
- Open WebUI is the ChatGPT-style front end you actually use
- An OpenAI-compatible API at
/v1, so existing code can point at it - Offshore, no KYC, crypto billing, and nothing logged by a third party
Get it running
Size the server for the model
This is the one catalog app that genuinely wants resources: around 4 GB of memory and 30 GB of disk for Ollama, since model weights are large. Pick an offshore VPS accordingly, and remember that CPU inference is slower than a GPU.
Deploy Ollama
Install Ollama from the catalog. It exposes an OpenAI-compatible API, which is what the front end and your own code will talk to.
Deploy Open WebUI and point it at Ollama
Install Open WebUI and set its backend URL to your Ollama deployment. It also speaks to any other OpenAI-compatible endpoint, so you can mix a local model with a remote one.
Register immediately
Open the interface and create your account right away, because the first user to register becomes the administrator.
Ollama has no authentication of its own. Published on a public hostname, its API is an open endpoint: anyone who finds it can run inference on your server, burn your CPU and use your models. Keep it reachable only from your Open WebUI deployment or your own application, or publish it as an onion service instead. Treat a public Ollama URL as a misconfiguration, not a feature.
Open WebUI hands administrator rights to whoever registers first. If the URL is public and you have not registered, someone else can claim it. Register the moment it is up, then review who is allowed to sign up.
Anything you cannot paste into someone else’s service: client documents under NDA, source code, internal data, or research you would rather not have retained for training. The model is not the smartest available, but the data stays yours.
Without a GPU, expect useful but unhurried responses on smaller models. Size the model to the server rather than the other way around, and test before you build a workflow on top of it.
Start now
Spin up an offshore VPS and install both from the catalog, or read the documentation.