Open-source agent nanoMuse runs across phone, desktop and browser with one account

What launched

An open-source project called nanoMuse has landed on Hacker News's front page, with a listing dated 7 October, pitching a personal AI agent that runs on the hardware you already own rather than inside one company's cloud. According to the project's GitHub README, the whole set is compact enough for one person to run and host: the Android app, the desktop app, the web console and the relay that joins them all sit in a single repository under GPL-3.0-or-later.

The agent is deliberately modelled on Meta's Muse, with its own name and appearance — a small dragon by default, replaceable by describing a new look that image and video models then draw and animate. The README stresses the project is independent and not endorsed by Meta, which owns the Muse trademark. The current release, per the changelog, is version 0.1.40 from 6 October.

What the agent can do

The README lists a Linux shell, a browser, MCP servers and skills as the agent's standard tools. A mode called Hands on extends that reach to the apps on your phone and the windows on your computer by reading and acting on their screens — a way to cover software that never had an API. On iOS this is not possible, the project says, because the system forbids it.

Safeguards are central to the design. The agent pauses for confirmation before deleting, sending or paying, and an approval can be remembered for a single time, the current chat or always. Passwords and verification codes stay with the user: logins and CAPTCHAs are handed back to the human, and the agent resumes when they are done.

The agent also persists between sessions. Goals are checked on a schedule, background routines continue when the app is not open, and it composes a morning feed. Its memory — what it is, what it knows about you and when it wakes — lives in Markdown files that the user can read and edit. It can also answer inside chat platforms including Telegram, Feishu, DingTalk and WeCom.

How the devices cooperate

Each device runs a full agent locally. The phone app embeds Alpine Linux under proot with its own shell and browser; the desktop app is a plugin of the DeepSeek Harness agent framework, with a Python runtime driving the hands. Signed-in devices meet on the relay and can hand work to one another, so a task started on the phone can execute on the PC, and a message beginning with @Mac routes to that machine. Approvals come back to whichever device is in your hand.

The privacy split, per the README: the text of conversations travels through the relay, but files and screenshots stay on the device where they were made.

Models, costs and self-hosting

Signing in grants a free allowance of model use on the community relay, which the developer pays for; once it is spent, users attach a key from one of eighteen providers — OpenAI, Gemini, DeepSeek, OpenRouter and Bailian among them — or sign in with ChatGPT, Claude or Kimi plans they already pay for. Providers without image or video models simply have those features switched off, and the app says so.

Nothing has to leave the house, the README argues: the same relay can run on your own server, with self-hosting billed as one VPS and about an hour, or no server at all if you use your own key. Builds exist for Android 8.0 and up, iPhone and iPad via TestFlight, macOS 12 and up, Windows 10 and up, Linux and Docker, with the usual rough edges — the Mac build is not notarised and needs a right-click to open the first time.

A comparison table in the repository positions nanoMuse between Meta's closed, cloud-VM-based Muse and OpenMinis, the single-device project whose phone app nanoMuse builds on. The desktop hands are a port of ByteDance's UI-TARS-desktop.

Why it matters

Consumer AI agents have so far been either closed cloud products or single-device experiments. nanoMuse's wager is that a self-hostable relay plus device-local agents can combine cross-device reach with local-first handling of files and screens. The limits are real — iOS blocks screen-driven control entirely, conversation text still transits the community relay unless you run your own, and a developer-funded free tier is not a durable cost model. But as a front-page release it is an installable, GPL-licensed alternative for anyone who wants an agent under their own roof, and its approval-first interaction pattern is a useful reference for agents that touch payments and deletions.