Watch OS
in progress· 2026
My own firmware for an ESP32-S3 smartwatch: press a button, speak, and the Nexus agent answers.
A Waveshare smartwatch with an ESP32-S3 and a 2-inch AMOLED screen. The idea is a watch that only acts as interface, microphone and speaker: transcription, model, tools and voice live on Nexus, and the watch is one of its voice channels.
Push-to-talk instead of a wake word: less battery and no accidental triggers. Audio goes back and forth in Opus over a WebSocket, so the answer can arrive while the model is still generating it.
How it is built
- The firmware restarted from scratch on Espressif’s ESP-Brookesia 0.8, which brings a multi-app launcher, settings, Wi-Fi and audio. I wrote the board definition myself: Brookesia does not include this model.
- The UI has its own design system: text on black, since a black pixel on AMOLED is off, thin lines and a single accent color. Inter with accented letters, a density of 2.0 like commercial smartwatches, and a boot screen with the name. It changes by overriding Brookesia’s JSON files, without touching its code.
- The development tools go over USB: screenshots of the pixels sent to the panel, simulated taps, CPU load and touch latency. They let the coding agent see and try the UI without me touching the watch.
Where it stands
The UI is done and needs a try on the wrist. Next comes power (sleep and wake from the button), then voice end to end with Nexus, then updates over Wi-Fi.
Before this
- 2026
Watch OS in C
The first version, in C on ESP-IDF: launcher, Wi-Fi, battery, time and settings. It had stopped before the voice app.

