KaozKit – JavaScript LLM agents on an engine built for microcontrollers

KaozKit is a Swift package that embeds the XS engine (Moddable) to run autonomous LLM agents written in JavaScript inside Swift applications. An agent is a small JS module that exports a run function, which can drive a language model, call tools, and read/write memory. The JS heap can be snapshotted to disk and restored in a fresh process, allowing resident agents to survive app restarts with their full state.

The package is organized into layered products: KaozJSCore, KaozJS, KaozHostC, KaozKit, KaozMLX, and a headless CLI called kaoz. It supports multiple LLM providers, including native Swift providers like Anthropic, Google, Ollama, OpenAI, and on-device Apple Intelligence, as well as JS-defined providers. Tools are registered in a ToolRegistry, with read tools safe by default and actuation tools opt-in and confined. Agents can schedule tasks, spawn sub-agents, and use persistent memory with embedding-based search.

KaozKit is designed for confinement and a small memory footprint, leveraging XS's embedded origins. The host.* surface is the only capability an agent has, and secrets never enter JS. The engine runs on a private thread with a clean C API, and await continuations settle correctly across the JS↔Swift boundary.