MimiClaw transforms ESP32-S3 microcontrollers into personal AI assistants by bridging Telegram messaging with Claude's LLM to enable conversational hardware control. The open source project follows a growing lineage of lightweight AI assistants including PicoClaw and the original OpenClaw, but marks the first implementation designed specifically for resource-constrained microcontrollers rather than Linux boards.
Written in C using the ESP-IDF 5.5 framework, MimiClaw requires an ESP32-S3 board with 16 MB flash and 8 MB PSRAM. Compatible boards include the LILYGO T7-S3, FireBeetle 2 ESP32-S3, ESP32-S3-DevKitC-1-N16R8, and Seeed Studio's XIAO ESP32S3 Plus. The system operates at just 0.5 watts while maintaining persistent memory across reboots and storing configuration in text files like SOUL.md for personality, USER.md for preferences, and MEMORY.md for long-term context.
Development requires a Linux workstation running the ESP-IDF 5.5 toolchain, which supports Ubuntu and Debian with standard package managers for installing git, cmake, ninja-build, and Python 3.9 or newer. Once compiled, the firmware runs bare-metal on the ESP32-S3 with no operating system, deliberately avoiding Linux on the target hardware to minimize power consumption and complexity. The dual-core architecture handles network I/O and AI processing on separate cores while supporting OTA firmware updates over WiFi.
The architecture routes user messages from Telegram through the ESP32-S3 over WiFi into an agent loop connecting to Claude's API. Beyond typical assistant tasks like reminders, MimiClaw's key capability is direct hardware control through natural language commands in Telegram. Users can read sensor data, toggle GPIO pins for lights or fans, and interact with connected actuators simply by chatting with the bot. Setup requires a Telegram bot token from @BotFather, an Anthropic API key, and optional Brave Search integration.
The project is available on GitHub under an MIT license with full build instructions for ESP-IDF environments. A waiting list at mimiclaw.io suggests preloaded boards may become available for purchase.
