MimiClaw
Run a complete OpenClaw-compatible AI agent on a $5 ESP32-S3 microcontroller chip. No Linux. No Node.js. No Mac mini. No Raspberry Pi. No VPS. Completely local-first memory, entirely portable, and exceptionally private.
The Ultimate Pocket AI Assistant
MimiClaw represents a complete rethinking of personal AI infrastructure. Hardware requirements for running complex agent loops normally demand expensive desktop computers, rented cloud servers, or heavy single-board computers. MimiClaw strips everything away, leaving only pure functionality embedded directly into an affordable microchip.
A community-driven software project, MimiClaw acts as an intermediary layer between your messaging applications and online language models. The software fits entirely on an ESP32-S3 microcontroller board with 16 megabytes of flash storage and 8 megabytes of PSRAM. These boards typically cost around five dollars and require less than half a watt of power to operate continuously.
You interact with MimiClaw through the Telegram messaging application. You type a message, and the ESP32-S3 chip receives it over a wireless network connection. The software then processes your message through an extensive agent loop. The language model analyzes the input, decides on actions to take, retrieves stored memories, and generates a response. After completing these steps, MimiClaw sends the reply back to your smartphone.
Because it operates on a microcontroller, MimiClaw removes the need for operating systems completely. It bypasses Linux, Windows, macOS, Node.js runtimes, Python interpreters, and virtual private servers. It executes instructions bare-metal on the processor, maximizing efficiency and minimizing software complexity.
Core Principles
- 1
Tiny Footprint
Written natively in C using the ESP-IDF framework. No bloatware allowed.
- 2
Handy Access
Message it directly from Telegram. The chip handles the rest of the conversation logic.
- 3
Loyal Memory
Learns from detailed memory files, remembering user preferences across complete power cycles.
- 4
Energetic Operation
Powered by a simple USB connection. It runs twenty four hours a day, seven days a week.
01Hardware and Performance
Operating an artificial intelligence agent normally implies massive server installations. MimiClaw inverts this requirement entirely, placing the cognitive routing center directly onto a piece of hardware smaller than a credit card.
Microcontroller Specifications
The minimum hardware requirement specifies an ESP32-S3 board with dual-core processors. It demands 16 megabytes of flash memory space and 8 megabytes of PSRAM. This specific hardware configuration provides enough operating room to hold the firmware, manage the wireless networking stack, and store files persistently.
Dual-Core Processing
MimiClaw distributes workloads efficiently across the ESP32-S3's dual processor cores. The first core dedicates itself to managing the wireless network connection and handling incoming network traffic. The second core processes the heavier task of operating the AI agent loop, string processing, and file system management.
Network Flexibility
The software handles complex networking environments natively. It supports Over-The-Air firmware flashing, allowing you to update the system without connecting a USB cable. It also includes an HTTP proxy feature, specifically implementing CONNECT tunnel support for networks that restrict direct outgoing connections to specific ports.
02Local-First Storage Design
Most consumer AI products store your conversational data on external company servers, presenting a significant privacy risk. MimiClaw alters this approach entirely. The software writes memory files as standard, readable text documents onto the local Serial Peripheral Interface Flash File System available on the microcontroller.
The agent parses these files directly during conversational exchanges. The system structures the data logically, separating short-term context from long-term memory blocks. Interaction histories persist flawlessly across sudden reboots or prolonged power outages. Because the data resides entirely on the chip, privacy remains entirely within your physical control.
You can physically examine these logs and modify them. If the assistant develops incorrect context or you want to dictate specific biographical rules, you can directly rewrite the memory files over a manual connection without interacting with complex database interfaces.
03The Autonomous Heartbeat
A personal assistant must act proactively, rather than acting only upon direct commands. MimiClaw implements an autonomous lifecycle that enables independent action processing.
ReAct Implementation
The core behavior relies on the ReAct methodology. When responding to a message, the language model determines if it needs additional information before finalizing a response. It can emit specific commands to trigger internal system functions.
MimiClaw grants the language model access to precise network functions. It supports checking the current internet time, compiling web searches through the Brave Search API, and directly modifying its own scheduling calendars. The loop continues iterating until the language model confirms task completion, finishing with a final message transmitted to the user.
Cron and Heartbeat Systems
A built-in cron scheduler allows the assistant to request future awakenings. The language model generates requests dictating tasks like generating summary reports every few hours or producing morning reminders at specific timestamps. These jobs persist in local storage, guaranteeing flawless execution continuity.
Furthermore, the heartbeat service periodically checks a specific task file maintained on the local storage partition. If incomplete task instructions appear inside this document, the system triggers the agent loop independently. The language model wakes up, analyzes the tasks, completes the work, and autonomously messages the user with status updates.
04Installation Instructions
Setting up MimiClaw requires compiling the source code using the official Espressif development framework. You must prepare the build environment correctly before flashing the software onto your microcontroller.
user@machine:~/mimiclaw$ ./scripts/setup_idf_ubuntu.sh
user@machine:~/mimiclaw$ nano main/mimi_secrets.h
user@machine:~/mimiclaw$ idf.py -p /dev/ttyACM0 flash monitor
Configuration Requirements
The software implements a two-layer configuration system. Fundamental build-time defaults operate from within the source code files. You maintain physical network credentials, API keys, and model selections in the secrets header document before compiling the firmware executable.
You require a specific set of parameters to function correctly. You must supply your local wireless network name and password for internet connectivity. You must acquire a Telegram bot token by interacting with BotFather on the Telegram application. Finally, you need a language model API key, typically sourced from Anthropic's console for Claude access or OpenAI's platform for GPT models.
Hardware Considerations
Connecting the microcontroller correctly involves identifying the appropriate port hardware. Development boards equipped with the ESP32-S3 processor typically contain two separate USB-Type C ports on the circuit board layout.
You must connect to the port labeled for native USB Serial and JTAG connections. Using the port intended for external UART bridges disrupts the flashing protocol and generates terminal monitoring failures. Always verify cable orientation and electrical port labeling before attempting to write code to the device logic.
Architectural Lineage
The project draws inspiration from multiple preceding experiments focused on personal AI autonomy. Understanding the differences highlights MimiClaw's distinctive position within the software ecosystem.
OpenClaw Ancestry
OpenClaw established the baseline methodology for extensive, autonomous digital personas. It requires substantial background horsepower to operate, frequently demanding deployment on full operating systems operating alongside numerous Node package modules. MimiClaw reimplements the fundamental agent loop structure, stripping away language dependencies to execute directly as pure binary code on embedded silicon processors.
PicoClaw Progression
PicoClaw reduced software requirements significantly, attempting to function within a ten-megabyte memory limit on minimal Linux architectures. MimiClaw abandons the operating system layer entirely. By compiling directly against the microcontroller's low-level framework, it achieves more complex cognitive routing interactions using less memory, completely eclipsing PicoClaw's processing constraints.
Important Inquiries
Deploy Autonomy Today
Constructed purely in C to operate continuously without expensive infrastructure. Adopt a highly optimized embedded assistant designed strictly for privacy.
Access the Repository