An AI agent body today can listen when you press a button, show its status and replies on a light or screen, take a photo, read air sensors and run commands on a computer. It can't move, pick things up or work without Wi-Fi and the cloud. This is an honest list of what a small AI agent body does, based on Meta's Muse Gadgets SDK and the boards hackshop supports.
What an AI agent body can do today
The thinking happens in the cloud. Every Muse board pairs with the Muse app, joins your Wi-Fi and holds an encrypted session to Muse. The board is the part that hears, shows and senses.
Listen when you press a button
Boards with a mic use push-to-talk. You hold a button, the board sends a voice note, and Muse transcribes it and answers. The M5Stack StickS3, AIPI Lite, Waveshare round AMOLED, SenseCAP Watcher and Home Assistant Voice PE all work this way. The SDK docs don't describe a wake word.
Show status, text and pictures
Even the simplest board shows what it is doing. The ESP32-C5 DevKitC-1 uses a single status light: breathing orange for setup, green for connected and blinking red for an error. Screen boards show an animated avatar and Muse's replies as captions, and many can display pictures Muse sends. The reTerminal e-paper displays show a still status screen and black-and-white or six-color images.
See with a camera
The SenseCAP Watcher is the camera board. Camera support is off by default. Once you enable it in the build, Muse can call camera.capture and get back a JPEG. A capture takes under a second, start-up included, and the camera is powered only while a capture or the live view runs.
Sense the air
The SenseCAP Indicator D1S and D1Pro have built-in CO2 and tVOC sensors, and temperature and humidity come from a Grove AHT20 sensor included in the box. Muse reads them all at once with sensors.read. This needs Seeed's stock RP2040 firmware left in place.
Act on computers and your network
On a Raspberry Pi or another Linux box, Muse can run shell commands, read and write files, and check uptime, load, memory, disk and temperature. Programs on that machine can message Muse too. On ESP32 boards with PSRAM, the home-network tunnel lets Muse reach devices you own that have a local HTTP API. You can also add your own commands, such as reading a sensor or switching a relay.
Can and cannot, at a glance
| Ability | What works today | Limits |
|---|---|---|
| Hear you | Push-to-talk voice notes | No wake word in the SDK docs |
| Answer you | Text captions, or the Muse app on the Voice PE | Spoken replies need your own text-to-speech API |
| Show things | Status lights, an LED ring, screens and e-paper | The color E1002 takes about 30 seconds per refresh |
| See | Watcher camera, when Muse asks for a photo | No photo attachments on voice notes |
| Sense | Air sensors on the Indicator D1S and D1Pro | Extras like the StickS3's motion sensor aren't used yet |
| Act | Linux commands, local HTTP devices, your own commands | Cannot move, roll, grab or lift |
| Work offline | No | Needs Wi-Fi and Muse's cloud |
What it cannot do yet
Move
None of the boards hackshop supports has wheels, motors or arms. They stay where you put them. You can wire a relay and add a command so Muse can switch something on or off, but that is a switch, not a robot. hackshop doesn't plan robot kits that move yet, so a rolling or walking agent is a fully custom project for now.
Work without Wi-Fi and the cloud
No network, no agent. ESP32-S3 boards need 2.4 GHz Wi-Fi, and the brain lives in Muse's cloud. Meta's terms also say the SDK and tokens "are not a supported product" and "may change, stop working, or be withdrawn at any time". Cloud devices do get switched off: Humane's AI Pin stopped answering queries when its service shut down on February 28, 2025. The difference here is that the firmware is open source and builds never turn on Secure Boot, so you can reflash the board for something else.
Answer instantly
Every reply makes a round trip to the cloud, and some parts are slow by design. The E1002 refresh takes about 30 seconds, custom commands time out after 30 seconds by default unless you raise the limit, and redrawing your avatar takes several minutes, mostly waiting for Muse. The SDK docs give no reply-time figures for voice, so test on your own network before you rely on it for anything urgent.
Speak out loud by default
Muse replies in text. For spoken answers, you hook a text-to-speech API into the firmware. The StickC Plus2 has only a buzzer, and the Voice PE shows replies in the Muse app.
Run for days on a small battery
Pocket boards have small cells: 250 mAh in the StickS3 and a 450 mAh module with the AIPI Lite. M5Stack advises keeping the StickS3 below 75% volume on battery to avoid reboots. E-paper is the exception, with Seeed advertising a 3-month battery life for its reTerminal displays.
Match the body to the job
- A face to talk to: a round screen like the Waveshare ESP32-S3-Touch-AMOLED-1.75C.
- Eyes on your desk: the SenseCAP Watcher with its camera turned on.
- Air you can ask about: the SenseCAP Indicator, in the D1S or D1Pro version.
- Hands on a computer: a Raspberry Pi with the Linux SDK. The guide to Muse on a Raspberry Pi covers the setup.
- A voice remote: the StickS3, at $21.50.
If you want help choosing, the Muse board picker quiz narrows it down in a few questions.
Questions people ask
Can an AI agent body move around?
Not with the boards hackshop supports today. None has wheels, motors or arms, and hackshop doesn't plan robot kits that move yet. You can add a relay command so Muse switches something on or off, which is useful but is not a robot.
Does a Muse gadget work without internet?
No. Every board joins your Wi-Fi and holds an encrypted session to Muse, and the agent's thinking happens in the cloud. Without a network, the board can't reach its agent.
Can my agent see through a camera?
Yes, with the SenseCAP Watcher. Turn on camera support in the build and Muse can request a photo, which comes back as a JPEG. Photos are not attached to voice notes.
Why does my gadget reply in text instead of speaking?
Muse sends replies as text, and screen boards show them as captions. Spoken replies need a text-to-speech service that you connect in the firmware.
Sources
- Muse ESP32 Device SDK README
- Muse ESP32 SDK AGENTS.md
- Muse ESP32 SDK supported devices, features and Watcher camera
- Muse Linux Device SDK README
- Muse Gadget SDK Token Terms of Use
- TechSpot: Humane shuts down AI Pin business
- M5Stack StickS3 product page
- AIPI Lite product page
- Seeed reTerminal E1001
- Espressif ESP32-S3
Prices and details checked October 5, 2026.