A new paper sketches nanoMuse, an open-source counterpart to Meta's Muse, the personal AI agent Meta introduced in September 2026.
The paper's authors say they worked out how Muse functions from Meta's public documentation and a copy of its production prompt, then used that to propose nanoMuse as a GPL-3.0 licensed alternative. The design calls for one agent that runs across every device a person owns, with "hands" that can act on a phone's screen and a computer's. Devices would share a single conversation over a relay that anyone could host themselves, rather than one company's cloud. Every action the agent takes would pass through a check the authors call a Sentinel, and memory would be stored as plain files a person can read instead of hidden in a vendor's database.
Muse treats the personal agent as tied to one account, one set of devices, and one country. nanoMuse's pitch is that the same category of software, something that acts on your behalf for weeks at a time, does not have to be closed or geographically locked to work. That is a meaningful reframe: it argues personal agents are a software architecture question, not just a bigger model.
Worth noting: the paper is upfront that nanoMuse's size and cost are estimates, not measurements, and that an evaluation suite plus an open model for the "hands" are listed as roadmap items, still to be built. A blueprint and a GPL license do not make a working agent. That part comes later, if it comes at all.