Run everything on your own computer

Sovereign mind. The most private agent is one whose model runs on your own hardware. With Ollama, nothing you say leaves your computer, and nothing depends on an account someone else can close.

You'll need a computer with enough memory for the model you choose. Small models (3B to 8B parameters) run on most laptops; larger ones want a GPU.

1. Install Ollama and pull a model

Install Ollama from ollama.com, then pull a model that supports tool calling, which the agent needs to act:

ollama pull llama3.2

2. Connect it to Suzent

1. Open Settings → Providers → Ollama and click FETCH.
2. Tick the model you pulled and click Save Changes.

3. Use it for everything

Open Settings → Model Roles and set Primary to your Ollama model. Leave Cheap empty, and it will use the same model; or pull a smaller one for background jobs and assign it to Cheap. See Model roles.

Memory search needs an embedding model. To keep that local too, pull one:

ollama pull nomic-embed-text

Add ollama/nomic-embed-text under Custom Models on the Ollama card, then assign it to the Embedding role. If you leave Embedding empty, Suzent still remembers facts about you, but memory search is turned off.

4. Mind the tools that reach the internet

The model is now local, but some tools still go online when the agent uses them, such as web search and the browser. If you want the agent fully offline, turn the Web group off in the chat's tool selector (see Tools).

Check it worked

Start a new chat, pick your Ollama model in the model selector, and ask something. If answers are slow, try a smaller model; if the agent never uses tools, try a model with stronger tool support.