OpenCharm

Two characters for eyes and one for a mouth, drawn on a black screen. That is enough for 22 moods across 13 agent states: listening while you hold the key, thinking when a turn starts, working during a tool call, needs you on an approval, done when a turn completes, stuck after three tool errors in a row, failed, learned when a skill is created, noted on a memory write, cost when a budget limit passes, asleep, and idle. Because each face is text, anyone can write a new one and share it.

Hold the key to talk and the microphone opens, then closes the moment you let go. Press it for everything else: wake the charm, stop it talking, dismiss a message, or answer no. Tap the face and it reacts. When the ring turns orange, holding means yes. On the desktop charm the key is Option plus Space on a Mac and Ctrl plus Alt plus Space on Windows, and you can remap it. There is no wake word and no gesture to learn.

charmd is a small daemon that runs next to your agent, on a laptop, a Mac mini or a server. It handles pairing and the PIN, speech on your machine or through your own OpenAI or ElevenLabs key, and the tools the charm offers to your agent. The charm itself holds no API keys and stores no PIN, only hashes. Pairing works one agent at a time, so the face always belongs to a single session.

The desktop charm puts the eyes on either side of the notch, with a panel that opens underneath for captions and questions. On Windows the same thing sits in a black pill at the top center of the screen. The app is about 4 MB and uses roughly 2% of one core at rest, which matters for something that never closes. The board runs the same OpenCharm OS, written in C++.

The first level needs no code. Who the charm is, what it knows and the skills it can run are text files in the workspace, and you can ask your coding agent to rewrite them. The second level adds MCP servers and exposes five built-in tools: show_face, ask, notify, say and set_look, so an agent can pick its own expression. The third level is the firmware, a C++ core built with ESP-IDF and tested in an emulator before it reaches the board.

Most AI agent hardware arrives sealed. This is a ready-made Waveshare ESP32-S3-Touch-AMOLED-2.16 with a 2.16 inch 480 by 480 AMOLED screen, two microphones, a speaker amp, a battery and a case, listed at $31.99. Back up the factory firmware, flash OpenCharm OS over USB, then pair it in about five minutes. A printed shell adds about $1 and a phone strap about $2, so the whole thing lands between $32 and $35.