Project notes · Built systems
Lumen
An experimental local macOS agent with a native interface and approval prompts for tool actions.
Problem & motivation
Local model tools need a usable desktop interface, visible task state, and a way to review proposed actions.
How it works
A Python backend connects a native macOS window to Ollama or LM Studio, with local model discovery, optional voice, a tool approval queue, and persisted task state.
Evidence & limits
This is an Apple Silicon desktop prototype, not browser-hosted inference. The website provides project information; running the application needs a local provider. Tasks are processed one at a time, and source builds are unsigned and unnotarized.
What I learned & next steps
I learned to discover installed models rather than assume one fixed model name. I would replace speech-recognition-based wake detection with a smaller dedicated wake-word model.
Source & setup
See the repository README for installation, commands, and the underlying evidence.