Meta Open Sources Muse Agent Code for DIY AI Gadgets

Meta has published the source code for Muse, the AI agent the company unveiled alongside a push into do-it-yourself hardware, giving developers a free hand to run the assistant on their own devices. The release turns Muse from a closed product demo into something builders can install on off-the-shelf components. Meta pitched the move as a way to let outside developers decide what an AI gadget should look like rather than shipping one fixed device.

The company outlined a handful of starter projects to show what the code can do. One suggestion is a color E Ink display that runs Muse as a low-power reminder board, refreshing only when the agent has something to surface. Another is an HDMI stick that pipes Muse onto a television so the agent appears on a large screen rather than a phone. Meta did not put a price or ship date on any of these builds, because none of them are Meta products — they are examples for people writing their own firmware.

Open sourcing the agent is an unusual step for a company that has spent the past two years competing on model quality and tightly controlled app experiences. By handing out the code, Meta lowers the cost of experimenting with an AI assistant and lets hobbyists and small teams test form factors that a large hardware program would never fund. The trade-off is that Meta gives up some control over how Muse behaves and where it runs, though the company still sets the underlying model.

The timing matters. AI hardware has struggled to find a mass audience, with several dedicated assistants failing to justify their price once buyers realized a phone could do most of the same work. Meta's bet is that the next useful AI device will not come from a single product team but from many small experiments, some of which will fail. Releasing the code is a cheap way to seed that pool of attempts without committing to manufacturing.

What the release does not settle is how capable Muse actually is on constrained hardware. E Ink panels and cheap HDMI sticks have far less memory and compute than the servers these agents normally run on, and Meta has not detailed the minimum specs or whether a network connection is required. Developers will also want to know what data leaves the device and what Meta retains. Watch for the first community builds in the coming weeks, and for whether Meta publishes performance numbers for the agent running locally.

For now, the project sits somewhere between a developer kit and a marketing exercise. Meta gets attention for Muse and a stream of free prototypes; builders get an agent they can wire into whatever hardware they already own. Whether that turns into a real ecosystem depends on how many of those prototypes survive past the demo video.