Reading note. This article draws on the Muse Gadgets announcement of October 2, 2026, on the GitHub repository published by Meta, and on the coverage available at the time of writing (TechCrunch, ai-tldr, Shopifreaks). The documentation does not state explicitly where the model runs: our reading, that the computation stays on Meta's servers, is inferred from the architecture described (pairing token, Muse app, active account). We present it as such.

In one sentence

On October 2, 2026, Meta released Muse Gadgets: two software development kits (SDKs) under the free Apache 2.0 license, which turn a small ESP32 electronics board or a Raspberry Pi into an object driven by Muse, its personal agent launched a month earlier. The gesture is real and useful: the code is open, the hardware costs next to nothing, and maker culture is invited to the table. But what is open is the object's body, not its brain: to work, every gadget must be paired with a Muse account and go through Meta's servers. We see it as good news for hobbyist electronics, and as an opportunity to recall what the frugal approach, for its part, seeks to keep at home.

1. What was announced

Let us set the scene. On September 8, 2026, Meta launched Muse, a personal agent able, according to the company, to book a trip, fill in a form or make purchases on the user's behalf. It is powered by Muse Spark, a closed model from Meta Superintelligence Labs, the group's artificial intelligence division. The launch began in the United States, restricted to adults, with capped free access and two subscriptions reportedly priced at $20 and $100 per month.

Less than a month later, Meta is opening a door to hardware. The facts, as published:

  • Two open SDKs. Firmware (the software embedded in a board) for ESP32, and an SDK for Linux, gathered in a GitHub repository under the Apache 2.0 license, with the exception of the in-house avatar and a few third-party components.
  • Common hardware. On the ESP32 side, around fifteen boards are reportedly supported, including Waveshare and M5Stack models, with the ESP32-C5 DevKitC-1 recommended. On the Linux side, the Raspberry Pi 3B+, 4, 5 and Zero 2 W, or any Linux machine.
  • Simple commands. According to the documentation relayed by ai-tldr, the Linux gadget exposes commands to the agent such as system.run (run a command), file.read, file.write and device.health, which the developer can extend.
  • An in-house object as bait. Meta built 5,000 units of a small box, Muse Home Link, powered over USB-C, which connects Muse to the home network to control speakers, televisions or printers. It is offered to active subscribers in the United States, one per person.

The examples showcased range from a color e-ink display to an HDMI stick plugged into a television. The tone, set by Nat Friedman, the lab's head of product, is that of the tinkerer: connecting Muse to "anything lying around on the workbench."

2. Two hardware families, two roles

To understand what Muse Gadgets enables, and what it does not, we need to distinguish the two platforms.

ESP32 — A microcontroller, meaning a very small computer on a single chip, made by the Chinese company Espressif. It costs a few euros, uses very little power, handles Wi-Fi and Bluetooth, and serves as the basis for a large share of cheap connected objects. In practice, it cannot run a large language model.

Raspberry Pi — A complete computer the size of a credit card, running Linux, sold for a few tens of euros. It can run small models locally, slowly but genuinely.

In the Muse Gadgets architecture, both play the same role: that of organs. The ESP32 provides a microphone, a speaker, a screen, sensors. The Raspberry Pi provides hands: launching a script, reading a file, controlling a home automation setup. The reasoning, however, happens elsewhere. Pairing goes through the Muse app in developer mode, with a token obtained on Meta's website, and requires an active account. Nothing in the published documentation describes an offline mode.

3. Opening the body, keeping the brain

A real gesture of openness

It would be unfair to reduce the announcement to a marketing operation. Publishing under Apache 2.0 means the code can be read, modified and reused, including commercially. A developer can study how Meta handles audio on a board costing a few euros, how pairing is secured, how commands are exposed, and reuse these building blocks elsewhere. For schools, fab labs and small workshops, this is high-quality teaching material.

It is also a notable change of strategy. After years of closed hardware (headsets, smart glasses), Meta is letting the community invent the forms of its assistant. The company implicitly acknowledges that it does not know which object will carry tomorrow's personal AI, and prefers to have thousands of tinkerers look for it on its behalf.

But a dependency by design

The central point is this: a Muse gadget is a terminal. Its intelligence is rented, account by account, from a single provider. Three consequences follow.

  • No network, no object. An internet outage, a service failure, an account suspension, and the object goes back to being a mute electronics board.
  • A restricted geography. Muse is launching in the United States first; the Home Link is only offered to American subscribers. A developer outside this perimeter is building for a service whose availability and timeline they do not control.
  • A sensitive entry point. The system.run command allows the agent to execute commands on the machine. This is precisely what makes the Raspberry Pi useful, and precisely what calls for caution: we are handing a remote service, driven by a language model, direct access to a computer on the home network.

The code is free; the intelligence remains rented.

We wrote it about Fable 5: a rented capability is a revocable capability. Muse Gadgets is no exception to the rule. It simply makes it more concrete, because it is now embodied in objects sitting on a shelf.

4. The frugal approach, by comparison

What we call the frugal approach is not a rejection of technology. It is a method: start from real constraints (energy, budget, connectivity, distance) and design tools that keep working when those constraints tighten.

Our project Sati is its most direct demonstration. It is an educational and cultural assistant designed to run autonomously on a Raspberry Pi. Same board as Muse Gadgets, opposite philosophy: the model runs on the device, the data stays in the room, and the object works without a subscription or a connection.

Let us put the two logics side by side:

  • Where the object thinks. Muse Gadgets: in Meta's data center. Frugal approach: on the board, with a small model, possibly quantized to fit in memory.
  • What you gain. Muse: a powerful, multimodal agent able to act on the web. Frugal: a more modest tool, but available, predictable and under control.
  • What you pay. Muse: an account, a subscription beyond the free quota, a dependency. Frugal: engineering time, and accepting limited capabilities.
  • What remains if the service disappears. Muse: the hardware and the SDK code. Frugal: the entire tool.

This comparison is not meant to name a winner. An agent that books a plane ticket needs the power of a large model and access to the web; an assistant that reads a tale in Creole to a class does not. The question is not "local or cloud," but which task deserves which level of dependency.

5. What tinkerers can take from it

For a developer who wants to make the most of Muse Gadgets without getting locked in, we suggest a few simple principles.

  1. Separate the interface from the brain. The ESP32 code published by Meta handles sound, the screen, the buttons. Nothing prevents you, thanks to the Apache license, from drawing on it for an object that talks to another engine, including a local model on a Raspberry Pi on the same network.
  2. Plan for degraded mode. If the remote service does not respond, the object should still provide a minimal service: display, record, respond to a handful of local commands.
  3. Restrict what the agent can do. If system.run is enabled, confine it to an allowlist of commands, on a dedicated machine, with no access to personal data.
  4. Keep your own tests. As with Jev, keeping a set of concrete cases lets you check that another engine does just as well, on the day you need to switch.

6. Signals to watch

  1. An offline mode. If Meta publishes a version of Muse Spark that can run locally, even a reduced one, the logic of the project would change in nature.
  2. Geographic expansion. The arrival of Muse and the Home Link outside the United States, notably in the European Union, will show whether the gadgets are a global ecosystem or an American laboratory.
  3. Security incidents. An agent that executes commands on home machines will be tested, sooner or later, by hidden prompt injection attacks. Meta's response to these first cases will be instructive.
  4. Independent forks. Versions of the SDK connected to other engines, open or local, would show that the community is taking ownership of the body without accepting the imposed brain.
  5. The business model. Will the free quota, generous at launch, tighten as the objects multiply? A gadget whose use becomes paid changes in nature.

7. A situated word

From Réunion Island, the announcement has a familiar taste. Objects that depend entirely on a distant server: we know what becomes of them when the submarine cable saturates, when a cyclone cuts the network, or when a service simply decides not to launch in our region. At 9,000 km from Silicon Valley, the question "what does this object do when it is on its own?" is not theoretical.

That is why we welcome Muse Gadgets with interest and with method. Interest, because thousands of people are going to learn to solder, to flash a board, to design an object that listens and responds, and that knowledge will circulate well beyond Meta. Method, because we will keep building tools that think on site, even if it means thinking more slowly.

The Raspberry Pi is the same on both sides. What changes is where we decide to put the intelligence. For an island, as for a classroom, an association or a small workshop, we believe the right answer often starts as close as possible: on the board, in the room, with what we have. 簡

Sources and further reading

This document is updated if new elements emerge. Last revision: 文 October 4, 2026.