Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

stackchan-mcp

Custom firmware and a Python gateway that let an agent move, see, hear and speak through M5Stack's StackChan desk robot.

Reviewed by ReviewRamble editors · Updated

Setup docs Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.8/5

A carefully engineered way to give an agent a small physical presence on your desk, with real limits on the servos and a network default you should tighten before you plug it in.

What is stackchan-mcp?

stackchan-mcp, from the GitHub user kisaragi-mochi, turns the official M5Stack StackChan kit (the 2025 Kickstarter version, built on an M5Stack CoreS3) into a body for an MCP client. It replaces the kit's factory firmware with a build of xiaozhi-esp32 and adds a Python gateway that is a stdio MCP server on one side and a WebSocket server for the robot on the other. The README lists 30 tools: head movement on two neck servos, photos from the 320×240 camera, touch reads, avatar faces, mouth and blink, the 12 base LEDs, speech, listening and a beat-following music mode.

The gateway is on PyPI as stackchan-mcp and is MIT-licensed; the firmware links a GPL-3.0 servo library, so the firmware build is effectively GPL-3.0. The repository has about 184 stars, released gateway v0.20.0 and firmware-v1.18.0 on 23 October 2026, and had commits in November 2026.

How to install stackchan-mcp

uv tool install stackchan-mcp

Is stackchan-mcp worth it?

The physical stakes here are low and the project treats them seriously anyway. The only moving parts are two SCS0009 neck servos. The gateway rejects move_head calls outside yaw -90 to 90 and pitch 5 to 85 degrees, M5Stack's recommended range, and the firmware clamps pitch to 0 to 88 at every servo write, so a call that skips the gateway still cannot drive the head past the limit validated on real hardware. The pose-following mode adds an angular-velocity clamp and caps commands at 20 per second. Everything else the agent controls is a screen, LEDs, a speaker and a microphone.

Those last two are the bigger consideration. listen records from the robot's microphone and transcribes it, by default with faster-whisper on your own machine and optionally with OpenAI's API; take_photo sends the camera frame to your model provider. The upgrade_firmware tool can reflash the robot over Wi-Fi, though the gateway first checks that the file is a StackChan image with a valid SHA-256 and the device keeps a rollback slot.

The catch: the gateway's robot-facing side listens on 0.0.0.0, port 8765 for the WebSocket and 8766 for photo uploads and firmware downloads, and if STACKCHAN_TOKEN is not set it logs a warning and accepts any client. Set the token on both the gateway and the robot before first use. There is no simulator; testing means the real kit, which is forgiving. Setup involves flashing with esptool over USB, configuring Wi-Fi through the robot's captive portal and, for speech, running VOICEVOX as a separate service.

Our take: for anyone who owns the kit, this is the most complete MCP route we have read, with a changelog that records hardware-validated fixes in unusual detail. It suits a hobbyist or researcher exploring embodied agents, not a classroom of unattended devices on a shared network.

Where it earns its place

  • A developer building a voice-first assistant prototype

    Give a coding agent a desk presence that speaks status updates and looks up when spoken to

    Run the gateway with local VOICEVOX and faster-whisper so audio stays on the machine, and let the agent call say and set_avatar when long jobs finish. Software costs nothing; model tokens are the bill.

  • An HCI researcher studying embodied agents

    Run sessions where an agent reacts to touch, sound and what it sees

    Enable the head-touch sensor, log stackchan/event notifications and pair take_photo with listen. Set STACKCHAN_TOKEN and keep the robot on an isolated network for participant sessions.

  • A maker who streams music sessions

    Have the robot nod and flash its LEDs in time with the room

    Start beat_mode_start with a motion intensity and colour; the gateway estimates BPM from the microphone and drives head sway and the 12 base LEDs without the model in the loop.

Before you commit

Setup
Flash merged-binary.bin from the latest firmware release with esptool.py over USB, join Wi-Fi through the captive portal, install the gateway with uv tool install stackchan-mcp, set STACKCHAN_TOKEN on both sides, then add stackchan-mcp to your client as a stdio server. Speech needs the tts extra and a VOICEVOX service.
Integrations
M5Stack StackChan kit (CoreS3, SCS0009 servos, GC0308 camera); VOICEVOX by default, with Irodori, ElevenLabs and Fish Audio TTS options; faster-whisper or OpenAI for speech-to-text; mDNS discovery; optional Tailscale Funnel or a Cloudflare relay for off-LAN use.
Data and privacy
Photos and transcripts go to your model provider through the client; audio is transcribed locally unless you pick the OpenAI engine. The gateway listens on 0.0.0.0:8765 and 8766 for the robot and accepts any client without STACKCHAN_TOKEN; its optional HTTP MCP endpoint defaults to 127.0.0.1.
Cost at scale
Free software. Cloud TTS or OpenAI transcription are billed by those providers if you choose them; the default local engines cost nothing beyond the hardware.
Limits
Only the official M5Stack kit is targeted; self-built StackChans need matching pins. mDNS discovery fails on networks that block multicast. About 80 open issues and pull requests.

What’s good

  • Gateway and firmware both clamp head angles
  • 30 tools across motion, camera, audio, face and LEDs
  • Local speech-to-text by default with faster-whisper
  • Over-the-air firmware updates with image checks and rollback

What’s not

  • Accepts any client on 0.0.0.0 until a token is set
  • Replaces the kit's factory firmware
  • Microphone and camera access by design
  • Only the official M5Stack kit is targeted

Best for

Owners of the M5Stack StackChan kit who want an agent with a face, a voice and a camera on their desk.

stackchan-mcp alternatives

stackchan-mcp: common questions

Is stackchan-mcp worth it?
A carefully engineered way to give an agent a small physical presence on your desk, with real limits on the servos and a network default you should tighten before you plug it in. We rate it 3.8 out of 5.
How much does stackchan-mcp cost?
Free and open source: the gateway is MIT, the firmware build effectively GPL-3.0. You need the M5Stack StackChan kit.
What are the downsides of stackchan-mcp?
Accepts any client on 0.0.0.0 until a token is set; replaces the kit's factory firmware; microphone and camera access by design; only the official M5Stack kit is targeted.
Who is stackchan-mcp best for?
Owners of the M5Stack StackChan kit who want an agent with a face, a voice and a camera on their desk.
How much does stackchan-mcp cost for a team or in production?
Free software. Cloud TTS or OpenAI transcription are billed by those providers if you choose them; the default local engines cost nothing beyond the hardware.
Is stackchan-mcp safe for company data?
Photos and transcripts go to your model provider through the client; audio is transcribed locally unless you pick the OpenAI engine. The gateway listens on 0.0.0.0:8765 and 8766 for the robot and accepts any client without STACKCHAN_TOKEN; its optional HTTP MCP endpoint defaults to 127.0.0.1.
What are the best stackchan-mcp alternatives?
The closest alternatives we have reviewed are ROS MCP Server and WLKATA MT4 control stack.

Answers checked Nov 16, 2026.