What happens when an AI coding agent moves off the desktop? This short AtlasRepo demo shows an open-source Apple Watch bridge controlling real Codex work on a Mac: selecting projects, opening chats, checking task status and usage limits, and sending a prompt from the wrist.
Watch the 48-second demo:
The watch is not doing the coding itself. It is a compact interface to a bridge service that reaches the developer machine where Codex is running. That distinction matters: a useful remote control needs clear task state, deliberate confirmation, and safe handling of authentication and network exposure.
Good wearable-agent interfaces should make it easy to answer three questions: What is running? What will this action change? Can I safely stop it? Tiny screens reward focused controls, not a miniature copy of a desktop UI.
Browse related tools and implementation evidence in the AtlasRepo open-source catalog.
Discussion: Which agent actions belong on a watch—status, approvals, voice prompts, alerts, or something else? Where would you draw the safety boundary?