Meta has not sold its tiny AI companion yet, but someone has already built one at home. The cost was not thousands of dollars, not a laboratory, and not years of coding experience. It was one ESP32 board costing 50 US dollars, plus AI help to write the code.
This story is interesting not because the result is perfect, but precisely because it is honest: it started out disappointing, slowly became fun, and still leaves many limitations. For anyone who likes IoT, ESP32, and AI agents, this is the most concrete example of what a home maker can already do in 2026, and what they cannot yet.
Quick summary
- A senior Verge editor built a Tamagotchi-style AI companion himself before Meta’s official product goes on sale.
- The board used is the Waveshare ESP32-S3 1.75C, a round 1.75-inch touchscreen, with a total build cost of 50 US dollars.
- He claims no coding experience, and used Claude Code to reprogram the device.
- Features that eventually worked include multiple avatars, a virtual-pet care minigame, conversation history, voice, remote camera from a phone, a pocket watch, and a video player.
- Meta’s official product itself, at the time the source article was written, had no announced price, firm date, or full specifications.
What Muse Charm is, and what Muse Gadget is
First, separate two things that are easy to confuse. The first is Muse Charm, Meta’s portable AI companion product planned for sale in the holiday season. It is small, can be carried like a Tamagotchi, and shows a cute AI face.
The second is Muse Gadget, the term for devices that use Meta’s open-source code and token to connect directly to the Muse AI cloud. The Waveshare board used in this experiment is mentioned as one of the early options Meta suggests for building your own Muse Gadget. So the device the editor built is not an official Meta product, but an experimental device based on a development kit.
This distinction matters. Do not equate a home-built result with a consumer product that has been tested, carries a warranty, and is ready for everyone.
The parts and the budget: an ESP32 for 50 US dollars
ESP32 is a cheap microcontroller popular in the maker world because it already has WiFi and Bluetooth, sips power, and is easy to program. In this case, the board used is the Waveshare ESP32-S3 1.75C, a round touchscreen pendant shaped like a metal pocket watch.
The 50 US dollar figure is the author’s build cost, not the official price of Muse Charm. Do not convert it into an official price in Indonesia, because shipping, taxes, availability, and board revisions can differ. What we can conclude is that the entry barrier is low: the hardware is affordable, and the biggest hurdles are no longer the board’s price, but time, care, and understanding of security.
How it was made, the simple version
The flow, in plain language, goes like this. First, the board is plugged into a PC. Second, Claude is pointed at Meta’s GitHub repository to use the Muse Gadget code and token. Third, the device is paired with the Muse app on a phone. Fourth, through several further Claude Code sessions, the device is reprogrammed bit by bit until it has the desired features.
The term for this way of working is vibe coding: writing programs by describing what you want in everyday language to an AI, which then composes the code. It feels magical for beginners, but it also carries risks we will discuss below: you can end up with code that runs, without really knowing whether that code is safe, tidy, or has hidden flaws.
For voice, this experiment used English AI voices from ElevenLabs’ free tier. For connectivity, the device used Bluetooth and the phone’s camera, and as a precaution it was placed on a guest WiFi network, not the home’s main network.
Features that eventually got added
The most fun part of this experiment is the list of features added through iteration, not in one shot.
First, avatars. The early version had only one built-in avatar that felt mostly silent. After reprogramming, the device could show many avatars inspired by those shown at Meta Connect.
Second, a Tamagotchi-style minigame. The virtual pet inside has needs for food, play, sleep, and bathing, with four levels of hunger, boredom, and dirtiness. Giving snacks can make a mess and increase dirtiness, and too much play can tire the pet out. These are small details, but details like these are exactly what make the device feel alive.
Third, conversation history. By swiping the screen, the user can read Muse’s full answers, not just fragments of words appearing a little at a time.
Fourth, voice. Answers are no longer just text on screen, but can be read aloud with an AI voice.
Fifth, remote camera. A button on the pendant can make the phone take a photo from a distance, with a high-quality live preview. There is also a lower-quality preview over Bluetooth, so it does not have to use the home WiFi. The phone’s front camera, which initially appeared flipped, was also fixed in a later coding session.
Sixth, pocket watch. The device can display an analogue clock with three hands that actually tell the time, and this feature is said to have been completed in about 15 minutes via Claude.
Seventh, video player. The device was loaded with clips found, downloaded, and trimmed from the Muse Charm introduction, which can then play on that small screen.
Eighth, a simple but important audio rule. At first, releasing the talk button too quickly sent almost-empty audio, confusing the AI. The fix was to only send audio longer than one second.
What was disappointing at first
Honestly, the early version was not immediately impressive. After installation, the device felt like it did nothing useful. The first utterance was treated more like an audio clip to transcribe than a command. Answers felt slow, were not read aloud, and only appeared a few words at a time, with no full message view except through the phone app.
The built-in avatar was also mostly silent. The device did not automatically discover smart home devices the way a smart speaker usually does. The sample phone-camera streaming app initially had no remote shutter button, making it less useful for photos.
Even after many fixes, lag still appeared. The author links the remaining lag to Meta’s servers, which are sometimes responsive and sometimes slow. This is an important reminder: IoT devices like this do not stand alone; their performance is partly determined by the cloud behind them.
Limitations, risks, and things not yet confirmed
This is the section to read before you try it yourself. First, this home-built device has no camera of its own and no 5G cellular radio. It is also large for a pendant, and nearly fell several times. The author himself concludes that the device has no more Muse functions than his phone, and in fact fewer without the extra code he wrote.
Second, security and privacy. The author admits he does not trust Meta or Muse with his email and home network, and mentions reports of a reviewer’s home address leaking. He also admits he does not know whether his vibe-coded code is good or has fatal flaws. This is not a new accusation, but a practical reminder: do not connect an experimental device to important accounts and your main network without isolation such as a guest network, and do not assume code that runs automatically means code that is safe.
Third, smart home control. In the source article, the ability to control a smart home is mentioned only in theory, not shown working. So do not write or claim as if this device can already control your smart home.
Fourth, Meta’s official product. Price, firm release date, full specifications, and whether the official product will be as freely reprogrammable as the development kit, are all unconfirmed from that source page. The author closes with a wait-and-see attitude: let us see when the product actually ships. We should take the same attitude in this article.
Who a project like this suits, and who it does not yet suit
It suits you if you enjoy experimenting with ESP32, are comfortable flashing firmware, understand network risks, and consider the learning process as fun as the result. It also suits those who want to understand how AI agents work on a physical device, not just in a chat app.
It does not yet suit you if you are looking for a finished, stable, warranted product that is safe for children to use unsupervised. It also does not suit those wanting a primary smart home device, because this device is not designed as a smart home hub, has no camera of its own, and its performance still depends on servers and connectivity.
If you are a total beginner, treat this project as an intermediate learning project, not a first project. Start with a board that has clear documentation, follow Meta’s official guide for Muse Gadget, and keep experiments separate from your important accounts and networks.
Final verdict
This experiment proves one big thing: in 2026, the line between AI device users and makers is getting thinner. With a 50 US dollar board and an AI that can write code, someone with no coding background can turn a touchscreen pendant into an AI companion you can talk to, care for like a virtual pet, use to take photos from afar, and use to tell the time.
But that proof comes with an equally large footnote: the early result was disappointing, performance still depends on servers, the code has not been verified as safe, and the official product itself may not be as open as this kit. So admire the innovation, but do not romanticise the process. If you are tempted to try it, try it with the expectations of an experiment, not of a finished product.
Quick Q&A
Is this an official Meta product? No. This is a self-built experiment based on the Muse Gadget kit, not the consumer version of Muse Charm.
How much does Muse Charm cost? Unconfirmed from this source. The 50 US dollar figure is the experiment’s build cost, not Meta’s official product price.
Can this device control a smart home? Not proven from this source. That ability is only mentioned in theory, not shown working.
Is it safe for beginners? It can be learned, but treat it as an experiment project. Use a guest network, do not connect important accounts, and do not assume code that runs is automatically safe.
Sources
The main illustration in this article is a generated illustration, not an actual photo of the device. This article is an original summary and analysis based on facts from the source above, not a copy of the source text. Prices, dates, and specifications of Meta’s official product that have not been announced are not written as facts in this article.
