Muse Gadgets: Meta's Open-Source AI Hardware Comes With Strings

What Meta released

On October 2, 2026, Meta published Muse Gadgets, a set of free software kits for building physical devices that talk to Muse, its personal AI agent. 1 Nat Friedman announced the project on X and described it as a side project. The code appeared on GitHub the same day under the Apache 2.0 license. 1 Friedman leads product at Meta Superintelligence Labs. 2

The release has two main parts. The first is on-device firmware for ESP32 microcontroller boards. The second is a Linux SDK aimed at Raspberry Pi computers or any Linux machine. 12 Meta pitches the idea simply: program an off-the-shelf ESP32 or set up a Pi, then wire Muse into displays, buttons, sensors, actuators and whatever else is lying around the workbench. 4 Meta also built its own reference device, the Muse Home Link, which is free for Muse subscribers. 12

What makers are meant to build

David Singleton posted about the launch the same afternoon. He listed several example uses:

• an e-ink morning briefing
• a push-to-talk pocket device
• a "sysadmin for your homelab" 1

Meta's project site adds other suggestions. One is a colour E Ink display that shows reminders. Another is an HDMI stick that puts the agent on a larger screen. 4

The coverage differs mainly in emphasis. Think Facility treats the release as a practical how-to and lists the requirements plainly. 1 1950.ai frames it as a strategic move. Instead of selling consumers one fixed AI device, Meta lets people invent their own ways of interacting with an agent outside phones, browsers and apps. 3 The Indian Express places it inside a wider hardware push. It notes that Meta appears to be exploring other Muse devices, including Muse Charm, a little-known Tamagotchi-like gadget. 4

Open code, closed brain

The part most worth examining is what "open source" covers here. Think Facility states that a working gadget needs an SDK token and the Muse app. 1 SecurityOnline describes the arrangement as moving "from building hardware to licensing the brain." 2 The firmware and SDK are permissively licensed, so anyone can fork, modify and ship them. The intelligence they connect to is not. It stays on Meta's servers, behind Meta's credentials, and in practice is tied to a Muse subscription if you want the official Home Link.

This is a familiar model, and it is not dishonest. Apache 2.0 covers the code Meta released and nothing more. But hobbyists should read "open-source Muse hardware" correctly. Your device is open. Its usefulness depends entirely on a proprietary service whose terms, pricing and availability Meta controls. If Meta changes token policies or retires the service, the open firmware becomes a shell with nothing to connect to.

SecurityOnline describes the benefit to Meta directly. The company can mine community creativity for ideas for its next official hardware. 2 It also gets a crowd of unpaid testers working out which interaction patterns suit an always-available agent, which is the experiment 1950.ai describes. 3 Calling it a side project keeps expectations low while still seeding an ecosystem.

The homelab question

Singleton's suggestion of a "sysadmin for your homelab" deserves more scrutiny than it got at launch. 1 Running the Linux SDK on a home server and letting a cloud-hosted agent act as a system administrator means letting a remote model issue commands on a machine you own. Nothing in the launch coverage describes how permissions, sandboxing or command confirmation work in that scenario. Builders should not assume such safeguards exist by default.

The sensible approach is the one you would take with any automation that has shell access:

• run it under an unprivileged account
• keep it away from root
• isolate it from sensitive data
• treat the SDK token like a password

An agent that can restart services can also break them. A compromised token on a device with broad privileges is a much bigger problem than a compromised token on an e-ink calendar.

The takeaway

Muse Gadgets is a smart, low-cost way for Meta to push Muse beyond phones and glasses, and it gives makers real, reusable code. 23 It is not an open AI platform, though. It is an open on-ramp to a closed one. For the e-ink briefing, that trade-off is trivial. For anything that touches your network's administrative controls, it is worth slowing down before handing over the keys.