Ollama runs models directly on your own machine — no API key, no account, no data ever leaving
your computer. It's the fully offline option, at the cost of needing more RAM and generally slower
responses than a cloud provider.

• 8GB+ RAM recommended (models are more likely to run acceptably; small models can work with less).
• Slower on older machines — a cloud provider (NVIDIA's free tier included) will feel snappier.

Start the app with the Ollama compose override instead of the default:

docker compose -f docker-compose.yml -f docker-compose.ollama.yml up -d

Then pick Ollama in the first-run setup wizard. No key field — the wizard detects whether
Ollama is actually reachable and running.

Every model picker in the app fetches GET /api/ollama/models, which calls Ollama's own
/api/tags endpoint, and lists whatever's actually pulled on your machine — not a fixed guess.
If nothing's pulled yet, the picker shows "No local models pulled yet — type a model id below" and
you can enter one directly (e.g. qwen2.5:3b, llama3.1, gemma2:9b); Ollama pulls it on
first use.

Background/system work (embeddings, free-tier fallbacks) defaults to qwen2.5:3b — the smallest
model the app's auto-pull setup guarantees will actually be present. Bigger models you've pulled
yourself (llama3.1, mistral, etc.) show up in the picker once Ollama reports them.

By default the app looks for Ollama at http://ollama:11434 (the bundled Docker container). If
you're running Ollama elsewhere, set OLLAMA_BASE_URL in your .env file — see
Environment variables.

The most common cause is a model that was requested but never actually pulled — see
Ollama not responding.

• NVIDIA (free tier) — the other free option, cloud-hosted, faster
• Docker install