nanoMuse brings open-source AI agent to every device

Overview

A new open-source initiative called nanoMuse is aiming to decentralize personal AI agents by making them portable across every device a user owns. Unlike proprietary cloud-based solutions, nanoMuse operates as a self-contained ecosystem that runs on local hardware, including Android phones, desktop computers, and web browsers.

The project positions itself as a direct alternative to models like Meta’s Muse and OpenMinis. It emphasizes privacy, user control, and the ability for devices to communicate with one another without routing sensitive data through third-party servers. The entire codebase is licensed under GPL-3.0-or-later, signaling a commitment to transparency and community-driven development.

Architecture and Availability

The nanoMuse repository contains the complete stack required to deploy a personal agent. This includes the Android application, a desktop client for macOS and Windows, a web console, and a relay server that synchronizes interactions across devices. Because the system is designed to be lightweight, it can be self-hosted by users who prefer to keep their data entirely within their own network.

Installation is straightforward for technical users, with pre-built binaries available for Android 8.0+, iOS via TestFlight, macOS 12+, Windows 10+, and Linux distributions. For those interested in hosting their own infrastructure, the project provides Docker support and shell scripts to spin up a personal relay locally.

nanoMuse is an open-source personal agent for every device you own. It does things instead of answering questions, keeps working while the app is closed, remembers you, and stops to ask before anything you could not undo.

Users can sign in with an email or a mainland-China phone number to access a free allowance of model usage on the community relay. If this allowance is exhausted, users can switch to their own API keys from a wide range of providers, including OpenAI, DeepSeek, Gemini, and OpenRouter. The project recently released version 0.1.40 Clear, which resolved several interface bugs on iOS and improved error handling across all clients.

Core Capabilities

What distinguishes nanoMuse from traditional chatbots is its agent-like behavior. Rather than simply generating text responses, the system is designed to execute tasks. It integrates with Linux shells, browsers, and Model Context Protocol (MCP) servers. Through a feature called "Hands," it can interact with the screens of phones and computers, allowing it to control applications that do not have public APIs.

The agent is also proactive and persistent. It can run routines in the background while the application is closed and check goals on a scheduled basis. For example, it might draft a morning news feed or execute a series of automated workflows without requiring constant user supervision.

Security and privacy are central to its design. Before performing irreversible actions such as deleting files, sending payments, or submitting forms, nanoMuse pauses to ask for confirmation. This checkpoint is remembered contextually—whether for a single interaction, a specific chat, or permanently. Sensitive information like passwords and CAPTCHA codes must always be entered by the human user, ensuring that the agent never holds credentials it should not.

The system is also highly social and interconnected. Conversations are synchronized across devices using a single account. A user can start a request on their phone and have it continue on their desktop. By prefixing a message with a device name, such as @Mac, users can explicitly route tasks to specific hardware. Approvals and notifications are then delivered back to the device currently in the user's hand.

Privacy and Data Sovereignty

One of the primary value propositions of nanoMuse is data sovereignty. In the traditional cloud-Virtual Machine (VM) model used by many competitors, user data often leaves the home network and travels to remote servers for processing. nanoMuse flips this model by keeping the agent local.

The most recent documentation highlights how data flows through the system. When signed in, devices meet on the relay, but only the text of conversations travels through it. Files, screenshots, and sensitive media remain stored on the local devices where they were created. This architecture ensures that even if the relay is compromised, the core personal data remains secure on the user's hardware.

The project explicitly contrasts its approach with Meta Muse and OpenMinis. While Meta Muse relies on a cloud VM per user and OpenMinis is limited to the phone it is installed on, nanoMuse allows one account to span all user-owned devices. This cross-device communication is facilitated by a hub that enables devices to ask each other for resources and approvals, creating a unified computing experience.

Every refusal from the relay is displayed as one plain sentence or a card on every client, ensuring transparency about why certain actions were blocked.

Furthermore, the agent's "memory" is not stored in a black-box database. What the agent knows about the user, its configuration, and its wake schedules are all stored in Markdown files. These files are readable and editable by the user, providing full visibility into how the AI perceives and stores information. This level of transparency is increasingly rare in an industry dominated by closed-source, proprietary algorithms.