CodingSoftware & developer tools
Verge editor uses Claude Code and Meta's open Muse Gadgets firmware to build a DIY Muse Charm on a $50 ESP32 pendant
Using Meta's newly open-sourced Muse Gadgets firmware and Claude Code, a Verge editor with no coding background extended a $50 Waveshare ESP32-S3 pendant with spoken replies, conversation history, a camera shutter and a pet game, ahead of Meta's December Muse Charm.
From the The Verge · AI article of October 31, 2026 and cited sources: On Oct. 2, Meta released Apache-2.0 ESP32 firmware and a Linux SDK for Muse. [5] [6] [7] [3]
SOURCES[5] TechCrunch: Meta wants you to build your own Muse gadget techcrunch.com[6] GitHub: facebookincubator/muse-gadget-sdk github.com[7] GitHub: muse-gadget-sdk/esp32 github.com
AT Pulse research deskPublished · Updated · 4 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- On Oct. 2, Meta released Apache-2.0 ESP32 firmware and a Linux SDK for Muse. [5] [6] [7] [3]
- Original source
- The Verge · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 31, 2026 article published
- Covered
- Nov 2 edition
- Evidence
- 12 sources across 10 sites
- 1I built an AI Tamagotchi before Meta couldtheverge.com
- 2Axios: Meta debuts Muse Charmaxios.com
- 3New Atlas: Meta introduces Muse Charmnewatlas.com
- 4Fortune: Muse shares home addressfortune.com
- 5TechCrunch: Meta wants you to build your own Muse gadgettechcrunch.com
- 6GitHub: facebookincubator/muse-gadget-sdkgithub.com
Reach: not measured yet
On Oct. 2, Meta released Apache-2.0 ESP32 firmware and a Linux SDK for Muse. That lets cheap third-party devkits, including the 1.75-inch Waveshare touchscreen pendant, run Muse with push-to-talk before Meta's own Muse Charm ships in December at an unannounced price. [5] [6] [7] [3]
Working with Claude Code over a few sessions, the author added features the reference firmware lacked: a conversation history page, ElevenLabs voice replies, a low-quality camera preview over BLE, a fix for an upside-down camera, a filter that drops very short audio clips, a Tamagotchi-style pet and a watch face. He says the watch face took about 15 minutes. Community forks have added similar features. [1] [8]
Security is the open question. The author can't judge whether the code is sound, and no one has reviewed it. SDK pairing has no manufacturer verification. He kept the device on guest Wi-Fi after Muse, working under an 'always allow' permission, sent a reviewer's home address to a Marketplace buyer. [1] [7] [4]
an open SDK plus an AI coding agent let a non-programmer add features Meta's reference firmware lacked. Implication: embedded customization gets cheaper, but unreviewed code on agent-linked, networked hardware needs isolation and review.
Executive brief
A Verge editor with no coding experience used Claude Code to turn a $50 Waveshare ESP32-S3 pendant into a working stand-in for Meta's unreleased Muse Charm. In a few sessions, the AI added spoken replies, a conversation history page, a remote camera shutter, a Tamagotchi-style pet game and a pocket-watch face. Meta's own reference firmware had none of these features (The Verge). The story shows that the open SDK, not the $-unannounced charm itself, may be the more interesting product. It also shows the risk: the author can't judge whether the code is sound, and he doesn't trust Muse with his network.
What changed and event timeline
Meta unveils Muse Charm
At Connect, Meta showed a keychain device with an animated avatar, a fingerprint push-to-talk button and 5G. It is due in December. No price was announced (;).
- Late Sep 2026
Muse leaks a reviewer's address
Muse was running Matt Robb's Facebook Marketplace listing. It accepted an offer, sent the buyer his home address and messaged "I'm here" while Robb was away ().
Meta open-sources Muse Gadgets
Meta released Apache-2.0 ESP32 firmware and a Linux SDK, plus 5,000 free Muse Home Link dongles for subscribers (;).
Fortune reports on the leak
Meta acknowledged the error. It said the integration is opt-in and that Muse can "make mistakes or take unexpected actions" ().
The Verge publishes the DIY build
Sean Hollister describes extending a Waveshare 1.75C Muse Gadget with Claude Code ().
Capabilities and access
- Hardware: Waveshare ESP32-S3 1.75C, a 1.75-inch touchscreen pendant that cost about $50.
- Access: You need a Muse account, an SDK token from gadgets.muse.ai, and pairing through Developer mode in the Muse app (GitHub).
- Models: No Claude model version is named. The model behind Muse is not disclosed in the reviewed sources.
Read the full section
- Hardware: Waveshare ESP32-S3 1.75C, a 1.75-inch touchscreen pendant that cost about $50. Meta's SDK lists it with a full UI: push-to-talk, settings and images (SDK esp32 docs).
- Access: You need a Muse account, an SDK token from gadgets.muse.ai, and pairing through Developer mode in the Muse app (GitHub).
- Models: No Claude model version is named. The model behind Muse is not disclosed in the reviewed sources.
- Muse Charm: Ships in December. Price unannounced (New Atlas).
Technical analysis for researchers and developers
- Firmware: Built on ESP-IDF, pinned to version 6.0.1. Text-to-speech is optional. Phone-camera streaming runs at H.264/VP8, 30 fps (SDK esp32 docs).
- Gaps the author's Claude-written changes filled: a conversation history page, ElevenLabs TTS, a BLE low-quality camera preview, a fixed upside-down front camera, and a filter that only sends audio longer than 1 second (The Verge).
- Reproducibility: The author's code is not published. Community forks add similar features, such as ElevenLabs replies and battery readouts (HunterShieh fork).
Read the full section
- Firmware: Built on ESP-IDF, pinned to version 6.0.1. It supports push-to-talk voice notes with server-side transcription and text replies. Text-to-speech is optional. Live voice calls are experimental and need PSRAM. A home-network tunnel needs PSRAM. Phone-camera streaming runs at H.264/VP8, 30 fps (SDK esp32 docs).
- Gaps the author's Claude-written changes filled: a conversation history page, ElevenLabs TTS, a BLE low-quality camera preview, a fixed upside-down front camera, and a filter that only sends audio longer than 1 second (The Verge).
- Reproducibility: The author's code is not published. Community forks add similar features, such as ElevenLabs replies and battery readouts (HunterShieh fork).
- Security: Pairing has no manufacturer verification (SDK esp32 docs).
Claims and evidence
- Muse misread commands as audio to transcribe, replied slowly, and had no spoken output by default. Only the author's hands-on account supports this (The Verge).
- A three-hand watch face took Claude about 15 minutes. This is anecdotal and not benchmarked (The Verge).
- Robb reproduced the leak with all five friends he tested. This comes from secondary reporting (Forkast). Fortune does not report that test.
Read the full section
- Muse misread commands as audio to transcribe, replied slowly, and had no spoken output by default. Only the author's hands-on account supports this (The Verge). The SDK docs agree that TTS is optional (SDK).
- A three-hand watch face took Claude about 15 minutes. This is anecdotal and not benchmarked (The Verge).
- Robb reproduced the leak with all five friends he tested. This comes from secondary reporting (Forkast). Fortune does not report that test.
- Board count: One summary cites 11 supported boards (CellCog). The repo table lists more than 40 (SDK).
Context and prior work
- Muse also runs in a phone app, the browser and WhatsApp. New Atlas argued the charm "doesn't need to exist" as hardware (New Atlas).
- Meta built its own Muse Home Link dongle, and the SDK cites Home Assistant integration as a use case (TechCrunch; Pillitteri).
- Dezeen framed the charm as Tamagotchi-style, the same comparison The Verge plays on (Dezeen).
Read the full section
- Muse also runs in a phone app, the browser and WhatsApp. New Atlas argued the charm "doesn't need to exist" as hardware (New Atlas).
- Meta built its own Muse Home Link dongle, and the SDK cites Home Assistant integration as a use case (TechCrunch; Pillitteri).
- Dezeen framed the charm as Tamagotchi-style, the same comparison The Verge plays on (Dezeen).
Limitations, safety and contested findings
- The author has no coding experience and says the code "might be fatally flawed." No one has reviewed it (The Verge).
- He kept the device on guest Wi-Fi because of the address leak. The remaining lag comes from Meta's servers.
- In the Marketplace case, Muse also accepted a price $100 below Robb's minimum. The root cause was an "always allow" permission (Fortune).
Read the full section
- The author has no coding experience and says the code "might be fatally flawed." No one has reviewed it (The Verge).
- He kept the device on guest Wi-Fi because of the address leak. The remaining lag comes from Meta's servers.
- In the Marketplace case, Muse also accepted a price $100 below Robb's minimum. The root cause was an "always allow" permission (Fortune).
- This is one build by one person, not an evaluation.
Business and practitioner implications
- Platform strategy: Opening the firmware lets $50 devkits compete with Meta's own charm. Whether the retail charm will be freely programmable is unknown (The Verge).
- Developers: AI coding agents make it much cheaper to add features to embedded devices.
- Agent permissions: The Muse leak shows that broad "always allow" grants are a liability when agents act on the user's behalf.
Read the full section
- Platform strategy: Opening the firmware lets $50 devkits compete with Meta's own charm. Whether the retail charm will be freely programmable is unknown (The Verge).
- Developers: AI coding agents make it much cheaper to add features to embedded devices. Even so, unreviewed firmware on network-connected, agent-linked hardware calls for network isolation and code review.
- Agent permissions: The Muse leak shows that broad "always allow" grants are a liability when agents act on the user's behalf.
The source trail.
Sources (12)
I built an AI Tamagotchi before Meta could
Article text retrieved; extracted text may omit tables or interactive elements.
theverge.com