nanoMuse: The GPL Answer to Meta's Muse Ships 1.0 and a Paper the Same Week
nanoMuse is an open-source personal agent built in the image of Meta's Muse, and this week it shipped both a 1.0.0 release, codenamed Keel, and a technical report on arXiv that hit 82 upvotes on the Hugging Face papers board. The pitch in the paper's own words: assistants from 2011 answered and waited, agents from 2023 did a task and stopped, and Muse in September 2026 showed an agent for one person, with accounts, devices, memory and a conversation that lasts, closed, in a vendor's cloud, in one country. nanoMuse is the open counterpart. One agent, every device you own, GPL-3.0, from authors Guangyi Liu, Yong Liu and Jiangning Zhang.
The whole set is in one repo: phone app, desktop app, web console and the relay that joins them. Builds ship for Android, iOS via TestFlight, macOS, Windows, Linux and Docker. Phone and desktop share one conversation over the relay, which you can self-host so nothing leaves your house, or use the community relay with a free model allowance the developer pays for until you bring your own key. The agent has a Linux shell, a browser, MCP servers and skills, and with "Hands" on it drives your phone's screen and your computer's windows for everything that never had an API. Say it on the phone, it runs on the PC. Every action passes through a Sentinel, and it stops before deleting, sending or paying, with the permission remembered once, per chat or always. Memory is files you can read.
The report does something unusual for a clone: it reads Muse's design from Meta's public record and a copy of its production prompt, marks each claim by source, and defines the personal agent through five questions and three horizons before describing the open build. Size and cost are given as estimates, and the roadmap names what is still missing, memory with provenance, an evaluation suite for the hands, and an open model trained for them.
The GitHub repo is at 344 stars, so this is early. But the structure is the point. Muse, Instinct, Manus Cue and Google's coworker agent all launched the personal-agent category inside a vendor cloud in the last month. nanoMuse is the first one you can read, run on your own server and point at whatever model you want, and it put its Sentinel-before-every-action design in writing. It will be the reference implementation people compare the closed ones against.
Links: github.com/nano-muse/nanoMuse, arxiv.org/abs/2610.08699, demo.nanomuse.dev
← Back to all articles
The whole set is in one repo: phone app, desktop app, web console and the relay that joins them. Builds ship for Android, iOS via TestFlight, macOS, Windows, Linux and Docker. Phone and desktop share one conversation over the relay, which you can self-host so nothing leaves your house, or use the community relay with a free model allowance the developer pays for until you bring your own key. The agent has a Linux shell, a browser, MCP servers and skills, and with "Hands" on it drives your phone's screen and your computer's windows for everything that never had an API. Say it on the phone, it runs on the PC. Every action passes through a Sentinel, and it stops before deleting, sending or paying, with the permission remembered once, per chat or always. Memory is files you can read.
The report does something unusual for a clone: it reads Muse's design from Meta's public record and a copy of its production prompt, marks each claim by source, and defines the personal agent through five questions and three horizons before describing the open build. Size and cost are given as estimates, and the roadmap names what is still missing, memory with provenance, an evaluation suite for the hands, and an open model trained for them.
The GitHub repo is at 344 stars, so this is early. But the structure is the point. Muse, Instinct, Manus Cue and Google's coworker agent all launched the personal-agent category inside a vendor cloud in the last month. nanoMuse is the first one you can read, run on your own server and point at whatever model you want, and it put its Sentinel-before-every-action design in writing. It will be the reference implementation people compare the closed ones against.
Links: github.com/nano-muse/nanoMuse, arxiv.org/abs/2610.08699, demo.nanomuse.dev
Comments