Meta Releases Muse Gadgets: Open-Source SDK for ESP32 and Linux AI Hardware

Meta on Thursday released Muse Gadgets, an open-source firmware and software development kit that allows developers to build physical hardware connected to the company's Muse AI agent. The package — covering Espressif ESP32 microcontrollers and Linux-based computers like Raspberry Pi — ships under the Apache 2.0 license, meaning any manufacturer can build and ship Muse-compatible products without paying Meta a licensing fee or negotiating a partnership deal.

The release arrived 24 days after Meta launched the Muse agent itself on September 8, and one day after Wall Street analysts highlighted an agent-plus-hardware-plus-monetization stack as a structural risk facing established platform companies. Meta filled in the hardware layer of that stack by publishing it as free code.

Read more: Meta Muse Hits No. 1 on App Store, Then a Mac Zero-Day Hijacks Its Permissions

How Home Link Bridges the Cloud-to-LAN Gap

The most immediately usable component of Muse Gadgets is Muse Home Link, a USB-C powered dongle that Meta manufactured in a first run of 5,000 units, offered free to active US Muse subscribers while supplies last with shipping expected in October.

Home Link solves a structural problem that has blocked every cloud-based AI assistant from controlling household hardware. Muse lives on Meta's servers — each subscriber gets a dedicated virtual machine — but home devices sit behind routers on private local area networks that outside servers cannot reach directly. Previous approaches, including Amazon's method for building Alexa compatibility, required manufacturers to build their own cloud connections back to the assistant's servers, a process that took years of partnership negotiations and still left vast categories of hardware unconnected.

Home Link reverses the direction. The dongle, powered by an Espressif ESP32-C5 chip and plugged into a home's network, initiates an outbound encrypted connection to Meta's servers. Muse then traverses that tunnel inbound to reach the local network. Once inside, Muse can address any device that exposes a local HTTP or HTTPS interface — a smart TV, a network-attached storage device, a Home Assistant installation, a Sonos system — without the device manufacturer having written a single line of Muse-specific code.

Meta's SDK documentation states the implications directly: commands sent by Muse run with the permissions of the account that installed the software. If that account has administrator access on the host machine, Muse inherits those privileges. The documentation also notes that community-initiated device pairing — as opposed to manufacturer-certified pairing — carries no manufacturer verification and cannot prevent an active man-in-the-middle attack during the 10-minute pairing window. Meta recommends enabling flash encryption and notes that development signing keys are included in the repository; Secure Boot is not enabled by default.

Eleven Boards, a Linux SDK, and a Desktop Simulator

Beyond Home Link, the Muse Gadgets SDK supports direct developer builds targeting 11 ESP32 development boards confirmed at launch, including the ESP32-C5 DevKitC-1, Seeed SenseCAP Indicator, Home Assistant Voice Preview Edition, M5Stack StickS3, and several Waveshare AMOLED and e-paper display boards.

ESP32 is a Wi-Fi and Bluetooth-capable microcontroller family made by Shanghai-based Espressif Systems, which has reported cumulative shipments surpassing one billion units and sells approximately 200 million chips per year. The chip appears in commercial smart plugs, light bulbs, air quality sensors, and doorbells, and it is the most widely used connected microcontroller in the global maker and electronics hobbyist community. In April 2024, Espressif acquired a majority stake in M5Stack, a maker of compact ESP32-based development kits; two M5Stack boards appear on Muse's initial supported list, reflecting the existing supply-chain relationship.

The Linux SDK, published alongside the firmware, converts any Linux-capable device — Raspberry Pi models 3B+, 4, 5, and Zero 2W — into a Muse terminal in four commands. The Linux SDK exposes four capabilities to Muse: shell command execution, file read, file write, and device health reporting.

Meta added a desktop simulator that renders the Muse interface in a 412-by-412 pixel window, allowing developers to build and test UI layouts before committing to hardware. The SDK repository includes a documentation file named AGENTS.md, written specifically for AI coding agents rather than human readers. Nat Friedman, who leads product at Meta's Superintelligence Labs and announced the Muse Gadgets release alongside Alexandr Wang, described the intent: point a coding agent at the repository, and the agent can generate a working device build without requiring the developer to have prior embedded systems experience. Embedded development historically demanded knowledge of register-level programming, memory management, and hardware debugging; the AGENTS.md approach attempts to route around that barrier entirely.

Every device that pairs with Muse — whether built from the open SDK or assembled from a community design — requires a token issued by Meta through gadgets.muse.ai. The token is non-transferable and capped at 50 devices per credential. Hardware can proliferate under any manufacturer; the pairing credential stays within Meta's control.

Read more: Meta Confirms a Persistent Ubuntu Linux Cloud Machine for Every Muse User

The Android Playbook, Applied to the Physical Home

Meta's architecture for Muse Gadgets mirrors a structure that Google executed with Android beginning in 2008, now applied to the physical world rather than mobile handsets.

Under Android, Google published the Android Open Source Project under an open license. Phone manufacturers adopted it without paying per-unit fees. Competition among manufacturers drove handset hardware costs to commodity levels, expanding the addressable market. Google kept the services layer — search, the Play Store, account infrastructure — outside the open-source release, ensuring that no amount of handset proliferation transferred the monetization pipeline away from Google.

Muse Gadgets reproduces this structure for IoT hardware. The firmware and SDK are Apache 2.0, maximally permissive: any factory in Shenzhen, any startup making elder-care sensors, any hobbyist with a breadboard can build Muse-compatible hardware without a licensing agreement, without disclosing modifications, and without paying Meta. The more hardware exists that speaks the Muse protocol, the more households have a permanent connection to Meta's AI agent. Meanwhile, the element that creates the connection — the SDK token from gadgets.muse.ai — remains under Meta's exclusive issuance. Hardware commoditizes; the credential layer does not.

This structure reaches a market that IDC has measured at 892.3 million device shipments in 2024, forecast to grow 4.4 percent to 931 million units in 2025. Despite that installed base, no single cloud AI agent has established itself as a reliable task-completion layer across household hardware. The smart home remains fragmented into brand-specific applications: a thermostat app, a doorbell app, a speaker app, each with its own cloud and its own authentication. Muse Gadgets attempts to insert a single reasoning agent beneath all of them.

Where the Previous Generation of AI Hardware Failed

The economics of the Muse Gadgets design contrast sharply with the two most prominent prior attempts to put AI inference into consumer hardware.

Humane's AI Pin, a wearable AI device that retailed at $699, placed a language model and neural processing unit directly on the hardware. The device ran hot, responded slowly, and required its own cloud subscription on top of its purchase price. Humane sold its assets to HP for approximately $116 million in February 2025, a fraction of the capital it had raised. Rabbit's R1, a similarly positioned pocket AI device, did not achieve commercial traction.

Both products placed inference on or near the device. Meta's architecture places inference entirely in the cloud, on the dedicated virtual machine each Muse subscriber already has, and assigns the edge hardware only three responsibilities: sense the environment, carry out physical commands, and maintain the encrypted tunnel back to Muse. A capable ESP32-C5-based device costs a few dollars in volume. A Raspberry Pi Zero 2W retails for roughly $15. At those price points, the barrier to building a Muse-connected device approaches zero for an experienced maker and falls to "one weekend with a coding agent" for a developer willing to follow the AGENTS.md path.

The practical consequence is a radically different deployment model. OpenAI acquired the hardware design firm io — founded by Apple's former design chief Jony Ive — in a deal worth approximately $6.5 billion that closed in July 2025. That approach produces premium hardware designed and manufactured by a single organization, constrained in scale by one design team's bandwidth and one factory's throughput. Muse Gadgets delegates the manufacturing problem to the global supply chain while Meta retains the software and monetization layer.

Apple, Amazon, and the Competitive Stakes of a New Physical Entry Point

The timing of the Gadgets release is inseparable from a week of analyst attention on Meta's expanding reach as a structural threat to established platform companies.

On September 29, Bank of America analyst Wamsi Mohan warned that Meta's Muse agent "could pull online commerce activity away from the iPhone ecosystem," identifying Apple's risk as losing "product discovery, referrals and transactions" while retaining device sales. Apple shares fell approximately 2 percent that day to $331.08, and Meta's stock rose 0.4 percent to $718.53. Bank of America maintained its buy rating on Apple but said concerns were "overdone" even as it flagged the structural point: whoever owns the agent layer collects the routing economics that previously accrued to operating systems, search, and app stores. The note added that Apple's current Siri lacks Muse's capacity for background task execution and broad third-party service integration.

The following day, Meta completed the hardware layer of that stack by open-sourcing it.

Amazon blocked Muse from executing purchases on its retail platform on September 20, twelve days after Muse launched. That block applies to the digital commerce layer; it does not prevent Muse from reaching Amazon Echo devices or Fire TV units that expose local HTTP interfaces inside a home network. The Gadgets design means Muse can potentially address smart speakers, streaming devices, and connected appliances through the Home Link tunnel regardless of whether the device manufacturer has agreed to cooperate.

Apple HomeKit's certification-gated approach and Google Home's fragmented product history have left the smart home control layer effectively unoccupied by a capable general-purpose AI agent. Alexa, the most widely distributed voice assistant in household hardware, has not developed persistent task-completion and broad third-party service integration at the level that Muse demonstrated in its first weeks of operation.

Muse's Hardware Expansion in Context

Muse Gadgets is the third new surface for the Muse agent in three weeks, following a connector platform for merchants opened September 18 and a Connect conference on September 23 at which Meta previewed Muse coming to its AI glasses, a Mac version capable of controlling any application with user authorization, and a forthcoming pocket accessory called Muse Charm.

A Muse TV Stick — a plug-in HDMI dongle for televisions — is listed as coming soon on the Muse Gadgets website. Meta says it plans to release a Muse encrypted mode later in 2026 that will provide additional privacy protections for users who connect sensitive home devices to the agent.

The pattern that emerges across these announcements is a deliberate layering of Muse access points: phone, computer, eyewear, pocket companion, and now every household device that can be reached through an HTTP endpoint. Each layer provides Muse with additional context about the user's environment and habits. A Muse agent that knows a user's calendar from their phone, their work documents from their computer, their physical surroundings from their glasses, and the state of every room in their home from a network of Gadgets-connected sensors carries substantially more context into any task than a single-device assistant.

Meta CEO Mark Zuckerberg has described human-level AI as arriving within the next one to two years. Whether or not that timeline holds, the Muse Gadgets release reflects a specific architectural bet: that the decisive competitive advantage in consumer AI will belong to whoever occupies the most access points in a user's physical and digital environment before any single AI capability proves decisive. The hardware layer is being built from the bottom up, at commodity cost, by millions of developers who now have a free SDK and a coding-agent-readable manual to get started.