# NanoMuse, open source AI agent software for mobile and PC

> NanoMuse is an open source AI agent software with 56 points on Hacker News that acts on mobile and PC with editable memory and its own relay under GPL-3.0
> By José Maldonado · AI Agents · 2026-10-10

There is a scene that anyone who uses an assistant knows well. You ask it for something useful, it replies with a correct piece of text and that is where it ends, because the real work is done by you with your hands. That boundary between answering and acting is what NanoMuse wants to break, and Hacker News received it on October 7, 2026 with 56 points and about twenty comments, according to the original launch thread in the Show HN section.

Meta presented in September 2026 its personal agent [Muse](https://ai.meta.com/muse/), with a name of its own and persistent memory, but locked inside the manufacturer's cloud and available only in the United States. NanoMuse proposes the same idea with the pieces in plain sight, in a single public repository under the GPL-3.0 license with a first version dated September 25, 2026 and near-daily releases, according to its changelog.

The concrete promise is simple. A single agent with a name and a face that runs errands on your Android, iPhone, Windows, macOS, Linux or browser, that keeps working with scheduled routines even if you close the app and that asks for your approval before deleting, sending or paying. When you sign up you get a free allowance on the community relay, and when it runs out you connect your key from among eighteen providers, as explained in the [official NanoMuse docs](https://nanomuse.cn/docs/).

## What does this agent do when you ask it to run an errand?

Think of the difference between a search engine and a sharp intern. The search engine brings you links and you decide. The intern understands the assignment, opens the necessary applications and hands you the finished result. NanoMuse plays that second role, with a Linux shell, a browser and MCP servers, plus a Hands mode that operates apps without a public API through the screen, according to the [nano-muse repository on GitHub](https://github.com/nano-muse/nanoMuse).

The permission flow is designed so that the scare does not arrive by surprise. Before an irreversible action a stop appears with a configurable scope, once, for this chat or forever. You type the passwords and codes yourself. If a login or a CAPTCHA comes up, the agent hands it over to you and resumes when you are done.

That discipline separates a toy from a serious tool. Asking permission before what cannot be undone is the rule that lets you trust it with your accounts. And the system remembers your answer so as not to ask the same thing twice, which makes living with it surprisingly bearable.

### Collaborating with your digital environment

Another practical trick is that devices collaborate. You say it on the phone and the PC executes it, with an @Mac at the start of the message to direct the task to the right machine. Approvals come back to the device you are holding, so the moment to decide never catches you far away.

What is more, the agent does not shut down with you. It checks goals on a schedule, runs routines with the app closed and writes you a feed every morning with what it prepared overnight. It is like having someone who gets up early for you and leaves the summary on the table.

Memory also departs from the usual script. What the agent is, what it knows about you and when it should wake up are text files you can open, correct and version. If it says something about you that you do not like, you delete it from the file and it stops knowing it forever.

For the software's social life, it replies inside Telegram and several Chinese messaging apps widely used in offices. And it sports an avatar that you describe in words and that an image model draws and a video model sets in motion, with a small dragon as the default look.

## An agent without ties

Each device runs its own agent and they all meet in a relay that carries the text of the conversations. Files and screenshots stay where they were born, without traveling over the network. The Android phone runs inside the APK with a Linux Alpine under proot, and the computer uses the desktop app with the Python runtime for the hands.

That architecture, explained in the [NanoMuse paper on arXiv](https://arxiv.org/abs/2610.08699), allows a device to work without a server when needed. The underlying idea is that intelligence can be distributed and still remain a single agent with a single conversation in the user's eyes.

The relay deserves attention because it decides who is in charge of your data. If you use the community relay, the text of your chats travels through it with a free allowance funded by the developer. If you prefer not to depend on anyone, you set up your own relay on your server with a script and nothing leaves your home.

![A look at how Nano Muse works](https://content.observatorioblockchain.com/wp-content/uploads/2026/10/nanomuse-work.webp)
*A look at how Nano Muse works*

Models are interchangeable too. The catalog covers eighteen providers, among them OpenAI, Gemini, DeepSeek and OpenRouter, plus plans you may already pay for such as ChatGPT, Claude or Kimi. If a provider does not offer image or video, those functions switch off and the app says so clearly.

For anyone coming from the Linux world, the good news is desktop support with AppImage, deb package and tarball for x64, plus a Docker image for the relay and the web. You can try it on your distribution without wrestling with odd dependencies, something projects born only for macOS tend to neglect.

**Related article: [PowerShell on Windows, installation guide and practical uses](https://observatorioblockchain.com/en/technology/powershell-en-windows-guia-de-instalacion-y-usos-practicos/)**

## Competing with Meta

The comparison with Muse is inevitable because the authors themselves propose it. Their October 2026 report reviews five frantic weeks in the sector. Meta opened the season with Muse in early September and OpenAI closed it with its always-on dots at the end of the month. In between came Today, OpenMuse and Manus Cue. All share a scheme, one computer per person in the vendor's cloud with thin clients and a vendor model.

NanoMuse reverses that scheme on three points. First, the agent runs on your devices, not on a rented virtual machine. Second, it adds hands on the Android screen, something the cloud does not even attempt. Third, everything is open, including the relay, with the model of your choice. In exchange it acknowledges shortcomings, with no model of its own, no wallet and none of the polish of a large team.

That frankness works in its favor. The roadmap admits that the hands fail when pointing at the right element and when recovering from unexpected dialogs. And it proposes as a remedy an open evaluation suite with fixed tasks to give each release a score instead of an impression.

The difference matters for anyone who values digital sovereignty. A closed agent gives you convenience in exchange for your errands passing through someone else's computers in another country. An open, self-hostable one asks a little more configuration and gives you back full control of your data.

## Security explained as a nightclub bouncer

Imagine that the agent is a willing helper to whom you give the keys to your house. Without rules, it could throw papers in the trash or open the door to anyone. NanoMuse places a bouncer at every important door, the so-called Sentinel, through which every action passes. Deleting, sending and paying always ask for your approval.

The real credentials never reach the model and the secrets stay on your side. That strict separation greatly reduces the surface for damage if something goes wrong. The helper can be very smart without needing to know your passwords to do its daily work.

The design shows in everyday details. Each account sees only its own conversations on a shared device, and logging out asks what to do with them. Each relay refusal arrives as a simple sentence or a card in all clients, with no cryptic codes that nobody understands.

The Mac takes a real screenshot with ScreenCaptureKit or explains why it cannot, instead of inventing a substitute image that would confuse you. That kind of technical honesty is exactly what distinguishes projects that look after the user from those that only want to show off.

It is worth noting that total security does not exist in any software, and even less in a 0.1 version. Auditable code helps, but flaws remain in plain sight until they are patched. Try it first with secondary accounts and raise the stakes only when the bouncer proves it is doing its shift.

## What do you need to try it this very afternoon?

Installation covers almost anything you have at hand. On Android it requires version 8.0 or higher on arm64 via a signed APK always with the same key, which installs over the previous one. On iPhone and iPad it arrives via TestFlight while it goes through Apple's beta review.

For its part, on desktop there are builds for Apple Silicon and Intel from macOS 12, an executable for Windows 10 and the three Linux formats. The browser is the ideal entry point, with a demo on a simulated phone to browse before installing anything on your machines.

The first launch follows a friendly script. You open the app, sign in with email or phone and the agent already has a model to think with thanks to the relay allowance. The phone, the desktop and the web share an account and show the same conversations everywhere.

From the model settings you choose which key each task uses, from chat to images. There is an automatic mode that tries in order your chat provider, the NanoMuse cloud and the first key that works. And if a move fails, you can ask that specific turn to draw on the community relay.

Aspect

NanoMuse

Typical closed agent

Where it runs

Your devices, optional relay

Vendor's cloud

Code

Full GPL-3.0

Closed except for gadget SDK

Model

Your choice among 18 providers

The vendor's

Memory

Editable Markdown files

The service's black box

Permissions

Approval before the irreversible

Approval with third-party telemetry

Base price

Free allowance plus your key

Subscription or pay-per-use

**Related article: [Why Microsoft integrates Containers directly into Windows](https://observatorioblockchain.com/en/technology/por-que-microsoft-integra-containers-directamente-en-windows/)**

## What this launch says about the future of agents

The September flood leaves a clear reading. The industry already knows what it wants to build, an agent with a name, legible memory and messages that arrive on their own when they are worth it. The discussion now is where that agent lives and who controls it. NanoMuse bets on it living with you and the control being yours.

We will have to wait for the next versions to know whether the hands on the screen mature for everyday use, because today they are the greenest piece. The plan to publish an open test battery and an open model for those hands looks good. If the community responds, each release will bring a number instead of an impression.
