nanoMuse: An Open-Source Cross-Device AI Agent

nanoMuse is an open-source personal AI agent designed to operate across a user's entire device ecosystem, including mobile and desktop platforms. Unlike traditional chatbots, nanoMuse focuses on action-oriented tasks—performing operations within applications rather than simply answering questions—and maintains a persistent state across devices via a central relay.

Cross-Platform Availability and Installation

nanoMuse provides a unified experience across multiple operating systems, allowing a single account to synchronize conversations and tasks.

  • Mobile: Available as an APK for Android (8.0+, arm64) and via TestFlight for iPhone and iPad.
  • Desktop: Native installers are provided for macOS (12+, Apple Silicon and Intel), Windows (10+), and Linux (AppImage, .deb, and tar.gz).
  • Web: A simulated phone demo is available at demo.nanomuse.dev.
  • Self-Hosting: Users can deploy their own relay using Docker or a provided self-hosting script to ensure data remains within their own infrastructure.

Core Capabilities and "Hands-On" Automation

nanoMuse is designed to interact with software that lacks official APIs by using the device screen as a primary interface.

Action-Oriented Execution

The agent can utilize a Linux shell, a browser, and MCP (Model Context Protocol) servers. For applications without APIs, it uses "Hands" to interact with the screen, taking screenshots and performing UI actions to complete tasks.

Cross-Device Coordination

Tasks can be delegated across hardware. For example, a user can initiate a command on a mobile device and specify @Mac to execute the task on a connected computer. Approvals for these actions are routed back to the device currently in the user's hand.

Persistence and Memory

nanoMuse supports background operations, allowing it to run routines or check goals on a schedule even when the app is closed. User memory—including what the agent knows about the user and its wake-up triggers—is stored in editable Markdown files.

Safety and Human-in-the-Loop

To prevent irreversible errors, the agent is programmed to ask for permission before performing destructive actions, such as deleting data, sending payments, or sending messages. Users must manually enter passwords and CAPTCHAs; once completed, the user signals "Done" to resume the agent's process.

Technical Architecture

Each device runs its own local agent instance. The Android app utilizes Alpine Linux under proot to provide a shell, browser, and MCP environment. The desktop application is built as a plugin for the DeepSeek Harness, utilizing a Python runtime for UI interaction.

Devices synchronize via a relay that handles authentication and carries conversation text. To maintain privacy, files and screenshots remain on the local device where they were created and are not transmitted through the relay.

Comparison with Meta Muse and OpenMinis

Feature Meta Muse OpenMinis nanoMuse
Execution Environment Cloud VM per user Local phone Local phone, computer, or private server
Non-API App Support None (VM cannot touch local devices) Android accessibility CLI Screen-based interaction (Phone/PC); not supported on iOS
Device Interoperability Clients of one VM Single device Cross-device requests via hub
Model Access Meta proprietary Bring your own Relay allowance or user-provided keys
License Closed GPL-3.0 GPL-3.0-or-later

Model Integration and Customization

nanoMuse supports a wide array of LLM providers, including OpenAI, Gemini, DeepSeek, and OpenRouter, as well as existing paid plans for ChatGPT and Claude. The agent's visual identity is customizable; users can describe a look, and the agent uses image and video models to generate its own avatar (defaulting to a small dragon).

Community Feedback and Security Considerations

While the project has been praised for providing an open-source alternative to proprietary agents, community discussions on Hacker News highlight several concerns regarding security and privacy:

"audit the installer: a similar tool, state dir redirected, still wrote eleven other assistants' config files and a port 4400 daemon."

Users have expressed skepticism about running high-privilege AI agents directly on primary mobile devices, citing the risks associated with granting an agent the ability to interact with the screens and system functions of a daily-use phone.

Sources

Related

  • Project
    HKUDS/nanobot

    An ultra-lightweight, self-hosted personal AI agent framework that integrates multiple LLM providers, long-term memory, and various chat apps into a single runtime.

  • Project
    osaurus-ai/osaurus

    Osaurus is a native macOS app that lets you run private AI agents locally. Agents have their own memory, can execute code in an isolated Linux VM or macOS sandbox, call tools (file system, browser, Apple apps, image generation), and communicate securely via encrypted channels. The platform works with local models (MLX, Apple Foundation Models, Liquid AI) or cloud APIs (OpenAI, Anthropic, etc.), and includes privacy‑filtering, cryptographic identities, plugin support, voice input, and system‑wide shortcuts. Install via Homebrew or a .dmg; requires macOS 15.5+ on Apple Silicon.

  • Project
    liliMozi/openhanako

    A user-friendly AI agent desktop application that provides personalized assistants with memory, personality, and the ability to operate a computer and browse the web.

  • Project
    CopilotKit/openmuse

    A personal agent application with a persistent browser, Linux terminal, and Google integrations that allows users to delegate complex tasks and review agent actions.

  • Project
    GCWing/OpenBitFun

    OpenBitFun is a cross-device, highly customizable agent workspace that allows users to manage, extend, and run AI agents across desktop, mobile, and remote environments.