Inspiration
As AI coding agents such as Codex become capable of completing increasingly complex tasks independently, I found myself repeatedly switching back to my computer to check whether the agent was still working, had completed a task, encountered an error, or was waiting for my input.
I wanted a way to stay aware of my AI agent’s progress without constantly watching the terminal. That idea led me to build VibeStick: a physical desktop console for AI coding agents.
VibeStick brings agent status, usage limits, task alerts, and voice input out of the computer window and onto a small hardware device that remains visible on the desk.
What It Does
VibeStick is built on the M5Stack StickS3 and is currently designed primarily around Codex workflows.
It can display:
- The current status of Codex
- Whether a task is running, completed, waiting for input, or has encountered an error
- Codex 5-hour and 7-day usage limits
- Device information such as Wi-Fi status, battery level, and time
When the agent completes a task, encounters an error, or requires the user’s attention, VibeStick provides on-screen and audio alerts, reducing the need to repeatedly check the computer.
VibeStick also supports voice input. By holding down a physical button, the user can speak an instruction. The system transcribes the speech and automatically inserts the resulting text into the currently active application on the Mac.
How I Built It
VibeStick is a hardware and software system made up of embedded firmware, a local desktop service, and a macOS companion component.
The device runs on the M5Stack StickS3 and handles the display, physical buttons, speaker, network connection, and device status.
A local service on the computer reads Codex status and usage information, processes voice recordings, and synchronizes information with VibeStick over the local network.
The project also includes a macOS HUD that displays agent status and supports desktop interaction.
During development, I used Codex to help analyze the project architecture, implement features, troubleshoot communication between the hardware and the computer, write tests, and improve the installation process and project documentation.
Challenges and What I Learned
One of the biggest challenges was compressing complex agent activity into information that could be clearly understood on a small screen.
The device cannot display complete logs, so I had to determine which information matters most to the user: what the agent is currently doing, whether the user needs to take action, and whether the task has completed or failed.
Another challenge was keeping the device, the local desktop service, and the macOS application synchronized while handling issues such as network interruptions, device reconnection, recording failures, and system permissions.
Through this project, I learned that interactions with AI agents do not have to exist only inside chat windows or terminals. A physical device that remains visible without constantly demanding attention can also become a useful interface for working with AI agents.
Log in or sign up for Devpost to join the conversation.