Harness docs / Get started

Local model

Start any OpenAI-compatible server and the harness finds it: llama.cpp's llama-server (port 8080 by default), Ollama (11434) or LM Studio (1234). With one model it uses it; with several it asks once (above); /model list switches later. A server on another machine: set HARNESS_LOCAL_HOST. The service is built for mid-size models (Gemma 4 26B-A4B, Qwen 3.5/3.6 35B-A3B, GLM 4.7 Flash); much smaller models struggle.