Ollama + Liminal

When to use Ollama with Liminal

Choose Ollama when mail, code, or client data must not leave your network, or when you already run models on a GPU workstation. The agent loop, file tools, git, memory, and session JSONL still run locally; only the model weights execute on your hardware.

1. Start Ollama

Install from ollama.com and pull a code-capable model. Examples: qwen2.5-coder, deepseek-coder, or llama3.1. Tool calling quality varies by model; prefer models marketed for agents or coding.

2. Configure Liminal

Add to `.env` or Settings:

• Use the same slug as ollama list shows.
• Set AGENT_EMBED_MODEL empty if you want BM25-only memory without cloud embeddings.

AGENT_API_BASE_URL=http://127.0.0.1:11434/v1
AGENT_MODEL=qwen2.5-coder:latest
AGENT_FAST_MODEL=qwen2.5-coder:latest
AGENT_API_KEY=ollama

3. Desktop app path

Open Settings in the Liminal desktop app, paste the base URL and model slug, and send a smoke-test prompt. Integrations (Slack, Xero) still work; only LLM traffic stays on localhost.

Performance expectations

Local models are slower than cloud APIs on most laptops but improve with a discrete GPU. Long agent turns with many tool rounds need a model that follows tool schemas reliably; if loops stall, try a larger quant or switch the fast model to a cloud endpoint only.