Jarvis: a macOS voice assistant prototype
An alpha personal macOS assistant with a Swift host and rendered HUD; native code owns machine interaction.
in progress
An ambient assistant has to connect model reasoning to a real machine without letting an attractive interface become the owner of everything. Jarvis explores that boundary as a personal macOS application with a native host and a rendered HUD.
CTO to Human: keep authority near the operating system
The project is explicitly for one user on one machine and remains in alpha development. Its public code is a place to study the design, not a claim that the application is a finished general-purpose assistant.
A Swift host owns machine interaction and agent state. A webview provides the visual surface. Separating those responsibilities gives the project a clear place to reason about permissions, failures and the effects of a requested action.
Engineer to Engineer: a typed boundary around the view
The documented architecture combines native Swift packages with React and React Three Fiber in WKWebView. A protocol-versioned JSON bus is the defined Swift/JavaScript boundary. Packages separate the agent loop, voice, memory, tools, replay and boot-health behavior; repository boundary checks enforce parts of that separation.
Local memory uses SQLite search/vector components, while the model-provider interface includes a hosted provider and an Ollama fallback. The repository describes local voice and memory processing separately from hosted model inference. A local subsystem should not be mistaken for a claim that every turn stays offline.
The Jarvis repository records the architecture and ongoing typed-API migration. This journal follows the mechanisms that connect an expressive interface to a controlled native runtime.