How to set up a local coding agent on macOS

Sources: HN discussion
Summary: A widely discussed write-up walks through running a local coding agent on macOS with local model inference. HN commenters noted that llama.cpp can fetch models directly with -hf and LLAMA_CACHE without a separate downloader, pointed to ollama and opencode as an alternative stack, and reported mixed speedups from multi-token-prediction draft setups on Apple Silicon.
Why it matters: Local coding agents avoid per-token API cost and data exposure, and the discussion captures the current practical setup and its limits on Apple Silicon.