商业产业 Hacker News AI 2026-10-07T03:30 正文 ok 950 words

Show HN: NanoMuse —— 适用于手机与电脑的开源跨设备个人 AI Agent

英文标题:🔶 Show HN: NanoMuse – An open-source AI agent for your phone and computer

开源项目 nanoMuse 正式发布,旨在打造一款覆盖手机、电脑与 Web 端的开源跨设备个人 AI Agent。该系统不仅具备 Linux Shell、MCP 服务器与浏览器调用能力,还可通过屏幕视觉交互(Hands)操作无 API 的桌面与移动应用,并支持跨设备任务路由与后台自动化常驻运行。nanoMuse 提供了由开发者赞助的社区中继免费额度,同时支持接入 OpenAI、Claude、DeepSeek 等 18 家模型提供商 API 或本地私有化部署。该项目以 GPL-3.0-or-later 协议开源并同步公开了 arXiv 论文,为开发者探索端侧自主 Agent 与多设备协同落地提供了完整的全栈开源参考。
打开原文相关 5 · 机会 3
media3.1 KB media668.9 KB media1.4 KB media1.3 KB media2.2 KB media1.3 KB media1.3 KB media1.3 KB media1.3 KB media3.1 KB media299.0 KB media73.5 KB

中文

nanoMuse — 面向你所拥有的每一台设备的开源个人 Agent

English | 简体中文 | 繁體中文 | Español | Français | Bahasa Indonesia | 日本語 | 한국어 | Русский | Tiếng Việt

GitHub stars Downloads Test Suite Try it in the browser Website Paper on arXiv GPL-3.0-or-later Discord

nanoMuse 是一款面向你所拥有的每一台设备的开源个人 Agent。 这是一个拥有独立名字与外观的单一 Agent,风格类似于 Meta 的 Muse:它专注于动手做事而非单纯回答问题,在应用关闭时仍能保持工作,能够记住你的偏好,并在执行任何不可撤销的操作前主动暂停向你确认。

nano 意味着整套组件足够精简,便于自行运行与部署:手机端应用、桌面端应用、Web 控制台以及串联它们的 Relay(中继)均包含在本项目仓库中,遵循 GPL-3.0-or-later 许可。免费、开源、非营利 —— 让我们共同打造它。 登录后即可在社区中继上获取免费的模型使用额度 —— 费用由开发者承担;额度用尽后,你可以使用自己的 API Key。相同的中继也可以部署在你自有的服务器上,确保数据完全无需离开你的本地局域网。最新版本:0.1.40 Clear —— 版本发布说明 · 在浏览器中体验。

视频/嵌入内容:nanoMuse 介绍短片 (英文) · 视频/嵌入内容:nanoMuse 介绍短片 (中文) · nanomuse.cn

  • 2026-10-07 📄 我们的论文已在 arXiv 发布。
  • 2026-10-06 🚀 最新版本:0.1.40 Clear。
  • 2026-09-25 🎉 nanoMuse 正式发布。

每个版本详情:releases。

所有安装包均可从最新 Release下载;若你所在的网络访问 GitHub 较慢,也可以在 nanomuse.cn/dl 获取完全相同的文件。打开应用后,使用电子邮箱或中国大陆手机号登录,它便拥有了用于思考推理的模型。手机端、桌面端与 Web 端共享同一账户,并同步呈现相同的对话内容。

动手做事(Does things)
拥有 Linux Shell、浏览器、MCP 服务器和各种 Skills —— 并且在开启 Hands 功能后,能通过屏幕画面直接操作你手机上的应用以及电脑上的窗口,覆盖一切从未提供过 API 的场景。 | 行动前先询问(Asks first)
在执行删除、发送或支付前主动暂停确认,设置可针对单次、当前对话或永久记住;密码和验证码均由你亲自输入。遇到登录或 CAPTCHA 验证码时会移交给你接管,点击“Done(完成)”后恢复自动化执行。
---|---
触达你的其他设备(Reaches your other devices)
在手机上发出指令,可在 PC 上执行;在消息开头加上 @Mac … 即可将任务派发到对应设备。操作审批提示则会回传到你手头的设备上。 | 持续不间断运行(Keeps going)
支持按定时计划检查目标进度,在应用关闭状态下继续执行例行日常任务,每天早晨为你量身撰写定制动态简报。
记住你的偏好(Remembers you)
关于它是谁、它对你的了解以及何时唤醒的信息,均以 Markdown 文件的形式存储,你随时可以查阅和编辑。 | 入驻你的常用聊天应用(Lives in your chat apps)
支持直接在飞书、钉钉、企业微信和 Telegram 中回复你的消息。
独具一格的专属外观(A look of its own)
只需进行文字描述,你的图像模型便会绘制其外观,视频模型则能使其动态呈现。默认形象为一条小龙。 | 支持接入任意模型(Any model)
可使用中继提供的免费额度,接入 18 家服务商(阿里云百炼 Bailian、OpenRouter、OpenAI、Gemini、DeepSeek 等)的自有 API Key,或直接绑定你已付费的订阅服务:ChatGPT、Claude、Kimi。若所选服务商未提供图像或视频模型,应用将自动停用这两项功能并予以明确提示。

每台设备均运行独立的 Agent —— 手机端在 APK 内部运行(基于 proot 的 Alpine Linux、Shell、浏览器与 MCP),电脑端运行在 nanoMuse Desktop 内部(基于 DeepSeek Harness 以及用于 Hands 的 Python 运行时)。设备登录后,它们在中继上建立连接并能跨端互相调用;对话文本在中继中传输,生成的文件和屏幕截图则完全保留在本地设备上。

Android、桌面端(Mac、Windows、Linux)、iPhone 和 iPad 以及围绕统一账户的 Web 应用:中继协助设备登录并在它们之间传输对话

docs/every-device.md 详细介绍了各设备支持,docs/hub.md 解释了框架机制,docs/cloud.md 说明了中继原理,docs/privacy.md 则阐述了系统保留的数据边界。

与 Muse 和 OpenMinis 的对比

| Meta Muse | OpenMinis | nanoMuse
---|---|---|---
Agent 运行位置 | 每位用户独立的云端虚拟机(VM) | 所安装的手机本地 | 你的手机、电脑或自建服务器 —— 跨设备共享单一账户
无 API 应用的支持 | 无法触及;云端虚拟机从不接触你的本地设备 | 基于 Android 的无障碍辅助功能 CLI | 将屏幕作为手机与电脑上的“手”:截图交互、优先调用 API,遇到登录操作交由用户接管。iOS 不支持该特性 —— 系统权限限制不允许
其他设备协同 | 均作为单一 VM 的客户端 | 仅限安装该应用的单台设备 | 设备间可通过 Hub 互相请求协作,审批确认推送至你当前使用的设备
模型支持 | 仅限 Meta 自有模型 | 自带模型(BYO) | 支持中继免费额度,或使用自有 API Key
开源许可 | 闭源 | GPL-3.0 | 基于 OpenMinis 构建,采用 GPL-3.0-or-later 协议

nanomuse.cn/docs —— 涵盖各平台安装指南、支持设备一览、Hands 交互、连接器、记忆系统、自托管部署以及通信协议。文档源码位于 docs/;每个版本的具体变更记录见 CHANGELOG 与 docs/releases/。

一台 VPS,一小时内搞定部署:详见 docs/self-hosting.md。提供三种方案 —— 使用自有 API Key 则完全无需服务器;通过 scripts/self-host.sh 搭建属于你自己的中继;或者为 Web 应用运行专属的本地运行时。

欢迎将它用于真实工作任务,反馈遇到的 Bug,并挑选感兴趣的方向贡献代码:CONTRIBUTING.md 详细介绍了开发环境搭建与协作规范,AGENTS.md 规定了编码 Agent 在此代码库中遵循的准则,路线图 (roadmap) 指明了切入点。Issues · Discussions · Discord。

如果你觉得 nanoMuse 对你有所帮助,请引用我们的学术论文:

bibtex @misc{liu2026nanomuseopensourcepersonalagent, title={nanoMuse: An Open-Source Personal Agent for Every Device You Own}, author={Guangyi Liu and Yong Liu and Jiangning Zhang}, year={2026}, eprint={2610.08699}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2610.08699}, }

nanoMuse 建立在开源社区众多前辈的工作之上;详细条款请参见 THIRD_PARTY_NOTICES.md。

nanoMuse 是一个独立的社区项目,与 Meta Platforms, Inc. 无任何关联或官方背书;Muse 为 Meta 的注册商标。龙形形象为该项目原创设计。

本项目遵循 GPL-3.0-or-later 协议。手机端应用基于 OpenMinis 1.13 (GPL-3.0) 演进修改而来,相关修改自 2026-09-24 生效 —— 参见 NOTICE。Python 产品线的更早版本曾采用 MIT 协议(Git Tag 为 pre-openminis)。

English / 原文

nanoMuse — an open-source personal agent for every device you own

English | 简体中文 | 繁體中文 | Español | Français | Bahasa Indonesia | 日本語 | 한국어 | Русский | Tiếng Việt

GitHub stars Downloads Test Suite Try it in the browser Website Paper on arXiv GPL-3.0-or-later Discord

nanoMuse is an open-source personal agent for every device you own. One agent with a name and a look of its own, in the style of Meta's Muse: 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.

nano means the whole set, small enough to run and deploy yourself: the phone app, the desktop app, the web console and the relay that joins them are all in this repository, under GPL-3.0-or-later. Free, open source, non-profit — let's build it together. Sign in and you get a free allowance of model use on the community relay — the developer pays for it; when it is gone, use your own key. The same relay runs on a server of yours, so nothing has to leave your house. Latest: 0.1.40 Clear — release notes · try it in the browser.

nanomuse-film-en-web.mp4

English · 中文 · nanomuse.cn

  • 2026-10-07 📄 Our paper is available on arXiv.
  • 2026-10-06 🚀 Latest version: 0.1.40 Clear.
  • 2026-09-25 🎉 nanoMuse is released.

Every version: releases.

All downloads come from the latest release; the same files are on nanomuse.cn/dl when GitHub is slow where you are. Open the app, sign in with an e-mail or a mainland-China phone number, and it has a model to think with. The phone, the desktop and the web share one account and show the same conversations.

Does things.
A Linux shell, a browser, MCP servers and skills — and, with Hands on, the apps on your phone and the windows on your computer through their screens, for everything that never had an API. | Asks first.
A stop before deleting, sending or paying, remembered for once, this chat or always; passwords and codes are yours to type. A login or a CAPTCHA is handed to you, Done resumes.
---|---
Reaches your other devices.
Say it on the phone, it runs on your PC; @Mac … at the start of a message sends the task there. Approvals come back to the device in your hand. | Keeps going.
Goals checked on a schedule, routines that run while the app is closed, a feed written for you each morning.
Remembers you.
What it is, what it knows about you and when it wakes are Markdown files you can read and edit. | Lives in your chat apps.
It answers in 飞书, 钉钉, 企业微信 and Telegram.
A look of its own.
Describe one, your image model draws it, a video model makes it move. A small dragon by default. | Any model.
The relay's allowance, your own key at one of eighteen providers (Bailian, OpenRouter, OpenAI, Gemini, DeepSeek and more), or a plan you already pay for: ChatGPT, Claude, Kimi. A provider without picture or video models leaves those two off, and the app says so.

Each device runs its own agent — the phone inside the APK (Alpine Linux under proot, a shell, a browser, MCP), the computer inside nanoMuse Desktop (DeepSeek Harness with the Python runtime for the hands). Signed in, they meet on the relay and can ask each other for things; the text of the conversations travels through it, files and screenshots stay where they were made.

Android, the desktop (Mac, Windows, Linux), iPhone and iPad, and the web app around one account: the relay signs the devices in and carries the conversation between them

docs/every-device.md explains the devices, docs/hub.md the frames, docs/cloud.md the relay, docs/privacy.md what it keeps.

Compared with Muse and OpenMinis

| Meta Muse | OpenMinis | nanoMuse
---|---|---|---
Where the agent runs | A cloud VM per user | The phone it is installed on | Your phone, your computer, or a server of yours — one account across them
Apps without an API | Out of reach; the VM never touches your devices | An accessibility CLI on Android | The screen as a hand on the phone and the computer: screenshots, APIs first, you take over for logins. Not on iOS — the system does not allow it
Other devices | Clients of one VM | The one it is installed on | Devices ask each other for things over the hub, with approvals where you are
Models | Meta's | Bring your own | The relay's free allowance, or your own
Licence | Closed | GPL-3.0 | GPL-3.0-or-later, built on OpenMinis

nanomuse.cn/docs — install per platform, every device, hands, connectors, memory, self-hosting, the protocols. The sources are in docs/; what changed in each version is in the CHANGELOG and docs/releases/.

One VPS, one hour: docs/self-hosting.md. Three ways — no server at all with your own key, your own relay with scripts/self-host.sh, or a runtime of your own for the web app.

Use it for a real task, report what broke, then pick something focused: CONTRIBUTING.md has the setup and the conventions, AGENTS.md the rules a coding agent follows in this tree, and the roadmap says where to start. Issues · Discussions · Discord.

If you find nanoMuse useful, please cite the paper.

@misc{liu2026nanomuseopensourcepersonalagent,
      title={nanoMuse: An Open-Source Personal Agent for Every Device You Own}, 
      author={Guangyi Liu and Yong Liu and Jiangning Zhang},
      year={2026},
      eprint={2610.08699},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.08699}, 
}

nanoMuse stands on other people's work; THIRD_PARTY_NOTICES.md has the terms.

nanoMuse is an independent community project, not affiliated with or endorsed by Meta Platforms, Inc.; Muse is their trademark. The dragon is the project's own.

GPL-3.0-or-later. The phone app is based on OpenMinis 1.13 (GPL-3.0), modified since 2026-09-24 — see NOTICE. Earlier versions of the Python line were MIT (tag pre-openminis).