Muse Gadgets open source: Meta opens hardware, keeps AI closed

Meta has released Muse Gadgets, a project that lets hobbyists wire its Muse AI assistant into hardware they build themselves. The code is genuinely open. The intelligence behind it is not. That gap is the most important thing to understand about the launch, and it is easy to miss under the "open-source" label.

What Meta actually released

Muse Gadgets consists of firmware and software development kits published on GitHub under the permissive Apache 2.0 license, in the facebookincubator/muse-gadget-sdk repository 1. There are two variants. An ESP32 Device SDK targets inexpensive microcontrollers, and a Linux Device SDK targets a Raspberry Pi or any other Linux machine 1. Several outlets described the same basic split between ESP32 boards and Raspberry Pi-class hardware 24.

The workflow is simple. Builders flash the firmware onto an off-the-shelf board and pair it with the Muse app on a phone. Muse can then speak through a speaker, listen through a microphone, draw on a small display, read sensors and trigger actuators 1. Meta's own pitch describes the devices as things "you build yourself," connected to "displays, buttons, sensors, actuators, and whatever else you've got lying on your workbench" 4. Suggested builds include a colour E Ink panel that shows reminders and an HDMI stick that puts the agent on a larger screen 4. One early walkthrough frames the entry point as a roughly $15 circuit board turned into a talking companion 1.

Alongside the SDKs, Meta is reportedly handing out free Muse Home Link hardware to some Muse customers 2. Reports on the giveaway's scale are thin and somewhat garbled, so treat the details with caution.

What stays closed

The coverage diverges on emphasis. The hobbyist-oriented write-ups lean into the possibilities of the open SDKs 14. A more skeptical analysis lays out the limits plainly. The Muse app, the agent harness, the cloud virtual machine the agent runs in, and the underlying Muse Spark 1.3 model are all proprietary 3.

The model is the only layer where Meta has said anything about openness. Open weights for Spark are "on the roadmap," but Meta has given no date, no license and no download 3. Even if the weights do ship, the Muse agent itself would still run only in Meta's cloud 3.

The hands-on guide makes a related point. Flashing the firmware does not install Muse onto the board 1. In practice, the gadget acts as a peripheral. It handles input and output and passes them through the phone app to Meta's servers, where the actual reasoning happens.

Why Meta is doing this

The Decoder's summary offers a strategic angle: opening the hardware layer helps Meta learn which AI device form factors people actually want 2. That fits with reports that Meta is exploring its own Muse-centred hardware, including a Tamagotchi-like device called Muse Charm, about which little is known 4.

The combination looks deliberate. Letting thousands of makers prototype pucks, watches, wall displays and desk companions amounts to crowdsourced product research. Meta gives up nothing core in the process, because every one of those devices still depends on its app, its cloud and its model.

The reading

Muse Gadgets is a real and useful release for makers. Apache 2.0 is a permissive license, the SDKs cover the two most popular hobbyist platforms, and the barrier to entry is a cheap board. Builders who want to give an assistant a physical body will find a capable toolkit.

The "open-source" framing deserves precision, though. What is open is the shell: the firmware that lets a microcontroller talk to Meta's service. The agent, the runtime and the model are closed. Builders cannot inspect, self-host or swap any of them. If Meta changes terms, pricing or availability, every Muse gadget changes with it.

This is a familiar pattern in AI. Companies open the edges of a platform to grow adoption while keeping the valuable center proprietary. The open weights for Spark, if they arrive, would be a meaningful step, but still not a self-hostable agent 3.

Makers should go in clear-eyed. Muse Gadgets is an open client for a closed service, not an open AI assistant. Anyone who needs an agent they can read, run and host themselves will have to look elsewhere 3.