Ollama (offline)

Use Ollama when you want in-app agents without a cloud API key. The LAN-night model is qwen2.5 7B on a Linux mini-PC / i9 — not a 3B, not llama3.2, and not an RX 580 / Polaris box. Full pick list: Getting started with AI.

Models

• Recommended: qwen2.5 7B (qwen2.5:latest, ~4.68GB). Create / start / health / stop works on a CPU box (a few minutes on an i9).
• Not the party path: qwen2.5 3B can call tools in a pinch. Do not run the LAN night on it.

Connect in PlayOn (recommended)

1. Open Settings → In-app agents and choose Ollama (offline).
2. PlayOn probes the Base URL (default http://127.0.0.1:11434/v1).
3. If Ollama is not reachable on this Home host and Docker is available, click Install Ollama. That starts a playon-ollama container (ollama/ollama on port 11434).
4. If Docker is missing, install Docker first (Docker on nodes — Settings → Nodes → Install Docker on Linux Home, or Docker Desktop on Windows), or copy the shown official Ollama one-liner and install Ollama manually, then Recheck.
5. Use the Model chooser, or Pull qwen2.5:latest.
6. Save settings, then open Map and send a chat to verify.

One-click install only targets localhost on the control plane (Home) host. For Ollama on another LAN machine, set Base URL to that host (for example http://192.168.1.50:11434/v1) and install/pull models there yourself.

Manual install (optional)

If you prefer not to use Docker from Settings:

1. Install Ollama on the party box (or another machine on the LAN).
2. Pull the LAN-night model:

ollama pull qwen2.5:latest

3. Confirm the OpenAI-compatible endpoint answers on http://127.0.0.1:11434/v1 (default), then connect in Settings as above.

Notes

• Tool calling quality varies by local model — qwen2.5 7B is the one to run.
• High-impact tools still require host Confirm in the UI.
• MCP clients (Cursor, OpenClaw, Hermes, …) do not need Ollama; they use their own models and a PlayOn access token. See MCP.
• Contributor env shortcut: PLAYON_LLM_MODE=ollama — Home users should use Settings.