An agent that lives on your devices, not in someone's cloud VM

What it does nanoMuse is an open-source personal agent in the mold of Meta’s Muse: it does things rather than answering questions. Each device runs its own agent — the phone app embeds an Alpine Linux environment under proot with a shell, browser and MCP support; the desktop app runs a Python-based harness that can drive your screen. Signed in, the devices meet on a relay and can hand tasks to each other, so a request spoken on your phone can execute on your PC.

The interesting bit The safety model is unusually explicit: it stops before deleting, sending or paying, hands logins and CAPTCHAs back to you, and only resumes when you say done. Its memory is plain Markdown files you can read and edit — no opaque vector store. And the whole stack (phone, desktop, web console, relay) is in one repo, self-hostable so nothing leaves your house.

Key highlights

• Cross-device: Android, iPhone/iPad, Windows/macOS/Linux, and browser, one shared account and conversation history
• “Hands” mode drives apps without APIs via screenshots — on the phone and desktop, though iOS system rules forbid it
• Works while closed: scheduled goals, routines, and a morning feed
• Model-agnostic: free community-relay allowance, your own key across eighteen providers, or existing ChatGPT/Claude/Kimi plans
• Also answers in Telegram and the Chinese enterprise chat apps (飞书, 钉钉, 企业微信)

Caveats

• The free model allowance is paid for by the developer — fine for a demo, not a promise
• macOS builds aren’t notarised and Windows triggers the usual “run anyway” dance; iOS is TestFlight beta only
• Screen-driven automation is inherently fragile; the README is honest that it’s a fallback, APIs first

Verdict Worth a look if you want a Muse-style agent you can actually own and self-host. If you wanted a polished, cloud-slick product, this is a community project with community-project edges — but refreshingly candid ones.