I’m Sergey, the creator of S1-ADV, an experimental software environment for M5Stack’s Cardputer ADV. I wanted the handheld to be a useful keyboard-and-screen interface to OpenClaw while keeping small tools on the device.
The architecture has two separate parts:
- ESP32-S3 firmware: keyboard, display, audio I/O and local applications.
- A Node.js bridge beside OpenClaw: forwards the device conversation to the host-side agent. Model access, configured speech services and optional development/build tools stay off-device. There is no LLM or Linux running on the ESP32.
The interaction idea is that a feature request can become AI-assisted code, a build, review and an approved update. Those are separate steps: receiving an agent reply does not mean firmware has been built or installed. Firmware automation needs deliberate host-side configuration.
The current getting-started guide keeps the first test small: get OpenClaw answering on the host, configure the bridge and a device token, flash over USB, connect to Wi-Fi, then verify a real typed answer on the Cardputer. Use your own endpoint and credentials.
This is an experimental project and a report of my own setup, not a validated one-click kit. I have not established fresh-device reproduction here, measured reliability, or compatibility with other Cardputer models.
Walkthrough (real device footage plus explanatory animation): https://youtu.be/-eJ83SK7VHQ
M5Stack designed the hardware. The project builds on OmniBot by Naz Louis and contributors; component-specific licenses and notices remain in the repository. This write-up was prepared with AI assistance against the project documentation.
I’d appreciate feedback on the small-screen UX: how would you distinguish a local tool, a pending host request and a completed device update without making the interface confusing?