hotdog
A zero-dependency coding-agent harness for local, open-weight models, pointed at llama.cpp, llama-swap, vLLM or any OpenAI-compatible endpoint on the user's own hardware.
hotdog's entire runtime is plain TypeScript run directly by Bun, with an empty dependency list: no npm install, no node_modules, no build step and no postinstall hooks, so there is nothing to compromise in a supply-chain attack and pinning a git tag means knowing exactly what is running. It is used daily by its author on a 128GB desktop running local open-weight models through llama.cpp and llama-swap.
It requires setting an API key even for a purely local server, since an unauthenticated local inference endpoint is described as both a free API for anyone on the network and an open prompt-injection surface. Markers, chat-template control tokens and tool-call delimiters found in untrusted tool output are rewritten to per-session aliases, so nothing a tool reads back can forge a fake tool call or harness system message.
It keeps a small system prompt and lets the user write and compose their own system prompts as tool-set profiles, supports one-shot mode, and lets individual features be disabled or extended.