nanoMuse Launches Open-Source Cross-Device AI Agent

Zhejiang University researchers have released nanoMuse 1.0.1 Ballast, an open-source GPL-3.0 personal agent framework designed to maintain a single identity across Android, iOS, Windows, macOS, Linux, and web browsers. The architecture allows independent hardware runtimes to coordinate through a relay system under one shared account. Users can dispatch remote commands from a mobile phone directly to a desktop computer using syntax like @Mac, receiving confirmation prompts on their handset while execution runs elsewhere.

The platform employs custom environment strategies depending on the host device. On Android, nanoMuse packages Alpine Linux inside proot, delivering a user-space environment featuring a shell, browser access, and Model Context Protocol support without requiring root permissions. Desktop instances leverage the DeepSeek Harness with a Python runtime. Graphical interface automation across Android, Windows, macOS, and Linux relies on a port of ByteDance's UI-TARS-desktop framework to manipulate screen elements.

Safety controls are handled by a dedicated Sentinel component that intercepts tool invocations, requiring manual authorization before an agent can pay, delete data, or send communications. Text-based history synchronizes via the relay, but screenshots and files remain strictly on the local device that generated them. nanoMuse supports 18 model providers, including OpenAI, Gemini, DeepSeek, Claude, and Kimi, while providing a free community relay allowance for connected users.

This release provides practitioners with a modular, self-hosted option modeled after concepts like Meta's Muse personal agent. By pairing standard protocol interfaces with broad multi-provider compatibility, engineers can deploy persistent cross-device background workflows without tying their infrastructure to a single proprietary cloud vendor or exposing sensitive local files.