Your $20 AI Assistant: Running OpenClaw on MimiClaw
I showed hardware enthusiasts an AI assistant running on a small board instead of a full computer, with conversations handled through Telegram.
Community Meetup · Singapore · 14 April 2026
An Assistant on a Small Device
At OpenClaw Singapore's Hardware Edition, sponsored by SuperAI, I demonstrated MimiClaw managing conversations and tasks on a compact device. The board stores conversation information and coordinates the work, while an online AI service generates the responses. It is not an entirely offline assistant.
The talk follows the practical work of getting the board connected and responding, including the limits of a small device and the problems that appear when connections are unreliable. It shows what can move onto inexpensive hardware and what still depends on services elsewhere.
Running the Assistant on an ESP32 Board
Local Execution and Remote Inference
MimiClaw runs the core agent architecture in C on ESP32-S3 without Linux or Node.js. I used an ESP32-S3-DevKitC-1 N32R16V with 32 MB flash and 16 MB PSRAM, above the 16 MB flash and 8 MB PSRAM minimum. Model inference uses a remote API.
Buffers, Sessions, and Message Queues
PSRAM holds the large prompt, history, and tool output buffers needed while a task is running. Local session files retain conversations, while queues handle incoming Telegram messages and outgoing responses.
Flashing and Diagnosing Connections
The setup used ESP-IDF v6.0, WiFi credentials, a Telegram bot token, and a model API key. Changing compiled secrets required a clean rebuild before flashing and monitoring the serial port. USB versus UART, SSL certificates, unstable connections, and remote model latency were separate troubleshooting concerns: a running board alone did not establish that the full Telegram and model path was working.