nanoMuse is a GPL-3.0-or-later personal agent spanning Android, desktop and a web console. Its October 7 0.1.41 “Choice” release addresses a practical problem: adding your own model-provider key previously did not make it obvious which model would handle which task, so the community-hosted allowance could still be used unexpectedly.

  • What changed: A Models page now separates chat, screen operation (“hands”), picture generation and short clips. Each slot shows its provider and model; saving a key offers to assign it to supported jobs, but does not silently change existing choices. If your selected model fails, the app offers a one-turn retry through nanoMuse Cloud rather than silently switching.
  • Getting started: The GitHub release currently has an Android arm64 APK and a desktop harness plugin bundle; the project also documents a browser demo and a Docker/SQLite self-hosted relay. The latest release describes standalone desktop installers and an iOS TestFlight build, but those desktop installer files are not in the GitHub release assets checked here; check availability before planning an install.
  • Cost and data path: The source is free; the project’s community relay offers a developer-funded starting model allowance, then you need your own paid provider key or eligible existing plan. A provider key can send model requests directly from the device, but sign-in and device coordination still use the community relay unless you host your own. The project’s own-key guide lists provider capabilities and local-model options.
  • Limits: This is a preview, not a verified safe automation system. iOS cannot operate its own screen; clip generation through a personal key currently supports only Alibaba Bailian. Screen-control agents need unusually broad permissions, so test with a spare device or constrained account before trusting one with private data. No hands-on test was performed for this capture.

The HN discussion raises the security-boundary question but provides no independent audit of nanoMuse. That is a useful question to bring to an evaluation, not evidence that the project is malicious.