Interesting project. We built something in a similar space: ClawWatch runs an OpenClaw agent (NullClaw, 2.8 MB static ARM binary) directly on a Galaxy Watch, paired with offline speech recognition via Vosk and the watch's built-in TTS. The agent connects to an OpenClaw gateway for LLM inference, so the watch handles voice I/O while the heavy compute stays server-side.
Different approach from Owl's capture-and-process model, but the wearable-as-agent-interface idea is the same. Would be curious how Owl handles the latency between capture and LLM response.
Website: https://clawwatch.app
Repo: https://github.com/ThinkOffApp/ClawWatch