Ollama (local models)

Official docs

Agent rules

1. Default local stack — Ollama is the fastest path to local inference on Mac/Windows/Linux for harness dev loops.
2. Model pick — coding: qwen2.5-coder, deepseek-coder-v2, codellama; general: llama3.1, mistral. Match VRAM (8GB ≈ 7–8B Q4).
3. OpenAI-compatible — point tools at http://localhost:11434/v1 when the client supports base URL override.
4. Never for ship gates — local models OK for explore/draft; verification and customer copy stay on cloud S-tier models per MODEL-ROUTER.
5. Privacy — prefer Ollama for secrets-heavy repos when cloud upload is disallowed.

Common tasks