ESP-Claw is an AI agent framework enabling edge AI on IoT devices, shifting from cloud-centric models to local sensing, reasoning, decision-making, and execution. It targets engineers and researchers seeking autonomous, responsive, and private IoT solutions.

How It Works

This framework merges LLM-driven dynamic decisions with deterministic Lua scripting. Its event-driven architecture triggers agent loops via any event, processed by a three-tier engine (LLM, Lua, Router). Users define device behavior conversationally via IM, with generated Lua code persistable for offline use. A structured on-device memory system uses summary tags for efficient retrieval, allowing agents to learn and evolve.

Quick Start & Requirements

• Installation: Offers "Online Flashing" for browser-based setup and updates, or build-from-source via basic_demo and ESP-IDF.
• Prerequisites: Requires Espressif chips. Optimal AI features depend on powerful LLMs (e.g., GPT-5.4, Qwen3.5-plus).
• Links: Documentation and demo examples are available.

Highlighted Details

• Chat to Build: Define device logic and generate Lua scripts via IM chat (Telegram, WeChat), with persistable local rules.
• Millisecond-Level Response: Event-driven system prioritizes local Lua rules for real-time, offline execution, using LLMs for complex tasks.
• Local Memory System: On-device structured memory (profile, fact, event, rule) with tag-based retrieval enables agents to learn and proactively suggest automations.
• MCP Protocol: Devices act as MCP Servers/Clients, exposing capabilities as AI-native tools (verb-noun naming) for seamless integration.

Maintenance & Community

The project is under active development. Community engagement is encouraged via GitHub stars. Specific community channel links or roadmaps are not detailed.

Licensing & Compatibility

The specific open-source license is not explicitly stated in the README, requiring further investigation for commercial use or integration.

Limitations & Caveats

The framework is under active development. Advanced AI features are heavily dependent on external LLM services and specific powerful models for optimal performance. The absence of a clearly defined license is a notable caveat.