Ollama

Runs open-weight models (Llama, Qwen, Mistral, Gemma, ...) on CPU or GPU and exposes an OpenAI-compatible /v1 API, so it serves SHC's `llm` capability socket with NO external API key. Its normal mode is fully local: it pulls a model on first run and answers /v1/chat/completions off that. This is the keyless backend behind apps that consume `llm` (agent-canvas, librechat, metabase, ...).