Meta has released the code needed to build hardware around Muse, its personal AI agent, and put it on GitHub for anyone to use. The Muse Gadgets project, published on 2 October, contains firmware for ESP32 microcontroller boards and a Linux software development kit, which together let an off-the-shelf board with a screen, a microphone or a set of sensors act as a Muse endpoint.
The licence is the permissive part. The repository sits at facebookincubator/muse-gadget-sdk under Apache 2.0, with named exceptions for a handful of third-party components, which means a company can build on it commercially without asking Meta for anything.
Alongside it, Meta has made its own hardware. Home Link is a USB-C powered dongle that connects Muse to a home network so it can reach televisions, speakers and anything exposing an HTTPS interface, including locally hosted APIs. The company built 5,000 units and is giving them to Muse subscribers in the United States, one each, shipping this month while stocks last.
What Is Actually In The Release
The repository is organised into ESP32, Linux and skills directories. The ESP32 firmware targets cheap boards with optional screens, audio input and output and sensors attached, while the Linux kit turns a Raspberry Pi or any Linux machine into a gadget that can run custom commands. Meta ships example implementations for Home Assistant and for routine system administration tasks, along with an AGENTS.md file so coding agents can work on the repository directly.
Several third-party boards are called out in the documentation, among them a Waveshare ESP32-S3 touchscreen, the Seeed reTerminal E1002, M5Stack StickS3 and the Home Assistant Voice Preview Edition. Suggested builds run from colour e-ink displays to an HDMI stick that puts Muse on a television.
Pairing is where Meta's side of the arrangement shows. A gadget needs an SDK token and has to be paired through the Muse mobile app after switching on developer mode, at which point it appears with a MuseGadget prefix. Meta also labels the whole thing experimental and warns that builders risk bricked boards and voided warranties.
The Code Is Free, The Agent Is Not
Giving away client code while keeping the service is a familiar structure, and this is a clear example of it. Every gadget built with this kit needs a Muse account to do anything useful, because the intelligence runs on Meta's infrastructure rather than on the board. The firmware handles the screen, the microphone and the sensors; the agent, the model and the subscription stay with Meta.
That gets Meta into product categories it has no interest in manufacturing. A company making a kitchen display, a workshop panel or a reception kiosk can add Muse without a licensing negotiation, and Meta picks up another surface for an agent it launched three weeks ago.
Muse Is Very New
Muse arrived on 8 September, running on a Meta model called Muse Spark, with a free tier and subscriptions at $20 and $100 a month. It is available in the United States on iOS, Android and the web, each user's agent running in its own virtual machine with a monitoring layer called Sentinel that holds back actions needing approval.
The expansion since has been quick. Meta extended Muse to small businesses on 29 September with connections to Shopify, Stripe, QuickBooks, Slack, Notion and a dozen other services, a day after setting up a Meta Enterprise Platform division under Chirantan Desai, formerly chief executive of MongoDB. Hardware is the third front in a month-old product push.
What Device Makers Should Check Before Building
The licence genuinely permits commercial use, so the question is not whether a manufacturer may ship a Muse-enabled product. It is what that product depends on.
Three dependencies stand out. The agent is a Meta service, so pricing, terms and feature availability are Meta's to change. Muse is United States only at the moment, which rules out shipping into Europe or Asia on this basis today. And every unit sold needs a customer with a Muse account, which makes the subscription a condition of the hardware working rather than an upsell.
The connectivity model deserves attention too. Home Link speaks HTTPS, and the examples lean on Home Assistant, so this reaches devices already exposing a network interface rather than the Zigbee, Thread and Matter estate that most smart home hardware actually uses. For industrial or commercial deployments, the same point applies to anything behind a gateway or a fieldbus, which still needs its own bridge in the way every device-level agent project eventually discovers.
Assistants Have Been Withdrawn Before
The risk in building someone else's agent into hardware is not hypothetical. Google retired Google Assistant on phones, tablets, watches and paired headphones on 4 September this year, replacing it with Gemini, and users cannot switch back. Meta itself discontinued Portal, its video calling hardware, in 2022.
None of that says Muse will go the same way. It does mean a product whose core function depends on a two-year-old service from a company that reorganises its AI operations regularly should have a plan for the service disappearing, which usually means designing the device to remain useful without it.
An Agent Looking For Somewhere To Live
What Meta has done here costs it very little and could pay off considerably. Publishing firmware under Apache 2.0 is cheap, the 5,000 dongles are a rounding error against Meta's AI spending, and the result is a route into hardware categories the company would never build itself.
For device makers, the calculation is different and worth doing slowly. Free code with a paid dependency attached is still a dependency, and the companies that built their products around voice assistants in the last cycle learned what happens when the platform owner changes direction. The integration work is cheap now. The question is what the device does on the day the service it relies on stops answering.
