Meta's Free Muse Gadget Kit Builds the Hardware, Keeps the Brain Locked Up

Meta released a free SDK on October 2 that lets hobbyists wire a microcontroller board or a spare Linux box into its Muse AI agent, and the access terms are worth reading closely before anyone orders parts. The kit, called Muse Gadgets, ships under an Apache 2.0 license and splits into two tracks: an ESP32 device SDK for small boards with screens, speakers or sensors, and a Linux device SDK that turns a Raspberry Pi into a Muse endpoint. Getting in requires an SDK token from gadgets.muse.ai, after which a developer points a coding agent at Meta's GitHub repository and builds against a short list of reference hardware: a Waveshare round AMOLED touchscreen, a Seeed color e-ink panel pitched for morning briefings and shopping lists, the M5Stack StickS3, an "AiPi Lite" desk companion, Home Assistant's Voice Preview Edition, and the Raspberry Pi 5.

Meta is also shipping its own first-party accessory alongside the SDK: Muse Home Link, a USB-C dongle that bridges the assistant to a home network so it can reach TVs, speakers and anything else answering to HTTP or HTTPS. The first run is capped at 5,000 units, free to Muse subscribers while supplies last, arriving within weeks. That is a pilot batch, not a retail rollout, and the gap between a 5,000-unit giveaway and the "build your own gadget" framing in Meta's announcement is the first thing worth separating out: most of what ships this month is still a test run.

The open license covers the hardware layer only. As remio.ai put it in a post examining the release, "Meta Muse Gadgets Are Open Source, but the Platform Is Not," the actual Muse agent, its underlying model, and the cloud backend that processes every request stay proprietary. A gadget built from the SDK is a peripheral that captures audio or sensor input and phones home to Meta's servers for anything resembling intelligence. That distinction mirrors a pattern Meta has leaned on elsewhere: it released a 30-billion-parameter model called Muse Glimmer under an Apache 2.0 license, with no monthly-active-user cap, days after launching the paid Muse Code product. Towards AI argued in "Meta spent three years preaching open source" that Glimmer is the distilled model, not the flagship, and that the giveaway does nothing to change Muse Code's $5, $15 and $50 monthly tiers.

The timing matters more than the license text. Chief AI officer Alexandr Wang promised a developer API for the Muse Spark model back in April, a promise that slipped repeatedly through the spring on cited bugs and infrastructure work, and then on no stated reason at all, according to coverage from The Next Web and Sherwood tracking the delay. Would-be hardware partners spent that stretch watching OpenAI and Anthropic sign up the API business Meta had promised. The open-source gadget SDK arrives roughly four months after that credibility gap, and it reads less like a sudden act of generosity and more like a way to generate builder enthusiasm while the underlying model access Meta actually controls remains the part it is least willing to hand over on predictable terms.

The release follows Meta Connect in September, where the company unveiled its first self-branded smart glasses with EssilorLuxottica, a Tamagotchi-style keychain called Muse Charm due in December, a roughly $1,300 VR headset, and a version 1.0 toolkit for building on the glasses hardware. Reality Labs has lost close to $88 billion since the metaverse push began, with $4 billion and $4.6 billion quarterly losses in the first two quarters of this year alone; CFO Susan Li has already signaled a shift in priority toward wearables and software, a shift that came with roughly 1,500 layoffs. Seeding an accessory ecosystem through free tooling costs Meta far less than building every form factor in-house, which is presumably the point.

It also sets Meta apart from how this category has gone for everyone else who tried to ship a single AI device. Humane's AI Pin raised about $230 million, shut down in February 2025, and had its remaining assets sold to HP for $116 million after reviewers and the company itself concluded the hardware was close to worthless. Rabbit's R1 is still being updated but was reported to have around 5,000 daily active users in recent coverage, kept alive through software rather than new units. OpenAI's hardware project with Jony Ive, the screenless device acquired through the io deal, is reportedly still aimed at a reveal in the second half of this year. Meta's bet is structurally different from all three: instead of betting one company's engineering budget on a single SKU, it is handing the hardware risk to whoever already owns a soldering iron and an ESP32 board, and keeping the part that generates recurring revenue, the model and the cloud it runs on, entirely to itself.