Argeu Thiesen has released Claudinho, a Claude Code plugin that materializes Anthropic's mascot, Clawd, onto physical ESP32 hardware. The project transforms the abstract status of an LLM session into a tangible desktop companion, reacting in real-time to the model's internal states such as thinking, tool usage, and errors. By leveraging existing Claude Code hooks, Claudinho provides a visual and auditory interface for monitoring long-running agent tasks.

Zero-Token Architecture and Local Connectivity

Crucially, Claudinho operates without external cloud services or Anthropic credentials. The plugin utilizes Claude Code's native hooks and status line features, which run locally in the background, to transmit short network notes to the ESP32. This architecture ensures that monitoring the LLM's activity consumes no additional tokens and does not degrade the model's performance. The board communicates directly with the local host over Wi-Fi, with the only external network request being a standard NTP sync for timekeeping.

Hardware Variants: E32R28T vs. Classic ESP32-C3

The project supports two distinct hardware configurations. The recommended E32R28T single-board computer features a built-in 2.8-inch touch screen and an integrated battery charger, simplifying assembly to just two soldered wires for power. This version supports optional voice synthesis via a small speaker and a battery for portable operation. The classic variant uses an ESP32-C3 paired with a Nextion 2.4-inch display, requiring four wires and a separate screen file upload. Both versions support 3D-printable cases available on MakerWorld.

Advanced Agent Visualization and Ticker

Claudinho extends beyond basic status icons to visualize complex agent behaviors. When Claude interacts with other models, such as OpenAI's Codex via the Codex plugin, the board displays a dedicated scene of Clawd and a cloud character conversing in binary speech bubbles. A scrolling ticker at the bottom of the screen displays plain-language descriptions of current tasks, such as "Compiling the firmware," pulled directly from Claude's existing command descriptions. This feature requires zero additional tokens as it repurposes text already generated for user approval prompts.

Key Takeaways

  • Claudinho uses local hooks and status lines, ensuring zero token consumption for hardware status updates.
  • The E32R28T board is recommended for new builds due to simplified wiring and built-in battery management.
  • The system visualizes multi-model interactions, including Claude-to-Codex and Claude-to-Gemini workflows.
  • Firmware updates are secured via SHA256 checksums and require a physical tap on the board to authorize.

The Bottom Line

Claudinho represents a clever engineering workaround that adds physical presence to LLM agents without incurring API costs. It proves that effective agent monitoring can be achieved entirely through local system hooks and lightweight network protocols.