Espressif Systems released ESP-Claw, an agent framework that runs the full sense-decide-act loop on ESP32 microcontrollers. The framework, which Espressif describes as a “chat coding” agent, handles tool execution, memory persistence, skill management, and peripheral control on-device. The LLM inference still happens on a separate server, but everything around it runs on the chip itself.

How It Works

ESP-Claw’s architecture mirrors OpenClaw’s: capabilities expose tools to the agent, skills define behavior through Markdown and Lua scripts, and a file-backed memory store (MEMORY.md) persists context across sessions. The agent reads its memory on every turn and can edit it through built-in tools, according to XDA Developers, which tested the framework on an Elecrow CrowPanel Advance ESP32-P4.

The built-in capability list includes Telegram messaging, Brave Search and Tavily web search, runtime Lua execution, file operations on FATFS, cron-style scheduling, I2C bus scanning, GPIO control, UART communication, and image analysis. Skills load as Markdown files paired with Lua scripts, meaning new behaviors can be defined through chat conversation rather than firmware recompilation.

XDA’s Adam Conway ran ESP-Claw on an ESP32-P4 connected to a self-hosted Qwen 3.6 27B model running on a Radeon RX 7900 XTX over LAN. Response times measured in seconds rather than the milliseconds typical of cloud inference, but Conway described the result as “one of the most interesting things I’ve ever put on an ESP32.”

The MCP Bridge

ESP-Claw speaks MCP (Model Context Protocol) both as a client and as a server. This means the microcontroller can expose its sensors and peripherals as tools to external agents, or consume tools from other MCP-compatible systems. For IoT deployments, this turns each ESP32 into a node that larger agent orchestrations can reach directly.

Practical Skill Examples

The framework’s Lua integration gives skills access to GPIO, MCPWM (motor control), LED strips, DHT temperature sensors, ADC inputs, SSD1306 OLEDs, and UART peripherals. A skill can be as simple as blinking an LED on a trigger or as complex as reading a temperature sensor every five minutes, logging to local storage, and alerting via Telegram if readings cross a threshold, according to XDA’s testing.

Skills survive reboots and can be created, activated, or deactivated through conversation. No recompilation required.

Agent Infrastructure at the Edge

ESP-Claw arrives as the agent compute stack fragments across tiers. Anthropic secured 300MW of SpaceX Colossus capacity for centralized inference. Google committed $200 billion to cloud TPU infrastructure. Datavault AI raised $60 million to build GPU edge networks across 100 U.S. cities. Now Espressif pushes the agent runtime to the smallest tier: microcontrollers running on local networks.

The question for builders is whether edge-deployed agents will remain isolated IoT controllers or become nodes in larger orchestration systems. ESP-Claw’s MCP support suggests Espressif is betting on the latter.