For people who can't use Meta Muse, they gave it three thousand stars: Zhejiang University's nanoMuse and "open-source alternative + paper + install-and-go"

nanoMuse is a personal agent open-sourced by Zhejiang University's APRIL lab. It can be installed on phone, computer and browser, and targets Meta's Muse, released in September. In 44 days after the repo was created it collected 3,221 stars and 388 forks. It took off thanks to three things stacked together: Muse was hot while Chinese users couldn't sign up, a paper that reached #3 on Hugging Face's daily papers chart, and an installer package that actually works. People who want to grow an open-source project's stars can learn a playbook here, but the "open-source Muse" position is already claimed, and the heat is receding.

A tweet: "No more worrying about Muse sign-up restrictions"

In Oct 2026 9, Chinese AI blogger @NFT_Chen on X posted:

Amazing! Zhejiang University's nanoMuse is a perfect open-source alternative to Meta Muse! A named, characterful personal agent that takes over all your devices! No more worrying about Muse sign-up restrictions!

The post got 35k views and 158 likes. The same week, @TylerBai_loveai (26k views), the Arabic-language @Sn0wbrave (14k) and @BrianRoemmele (10k) all reposted it. On Zhihu, questions about "an open-source alternative to Meta Muse" were opened several times a day, and column titles all read like "No need to scramble for Meta Muse" or "The Zhejiang University open-source version of the hot Muse is here". On Oct 13, Xinhua's client app also reported it.

The last line is the key. Meta launched Muse in Sep 8. In three weeks it hit 2.5M downloads and topped the US App Store free chart (per 什么值得买's industry coverage, unverified). Fortune reported over 10M users by Oct 22. But Chinese users face a signup wall. Meanwhile Hacker News has run bad Muse news for a month: "6.8GB filesystem export", "critical 0-day", "ignores user permissions", and a Techdirt piece calling it "a cute privacy and security dumpster fire" (392 points). Amazon even blocked its shopping feature as "unauthorized AI agents". Too risky to use, too useful to ignore. nanoMuse sits right between those two problems.

What it is

A personal agent with a name and a character (a little dragon by default), one account across all your devices:

• It can act: it has a Linux command line, a browser, MCP and skills. Once "Hands" is turned on, it can tap apps on your phone and windows on your computer, covering things that have no API.
• Cross-device: say something on your phone and it runs on your computer; start a message with "@Mac" to send the task to that machine. Requests that need approval come back to the device in your hands.
• Asks before acting: it stops and checks with you before deleting, sending messages or paying; you enter passwords and verification codes yourself.
• Remembers you: who it is and what it knows about you are stored in Markdown files you can open and edit.
• Bring your own model: logging in gives you free quota through a community relay (paid for by the developers); after that, switch to your own key. It supports 25 providers including Bailian, DeepSeek and OpenRouter, and you can reply through Feishu, DingTalk, WeCom and Telegram.

The install page is a table: browser demo, Android APK, iPhone via TestFlight, macOS / Windows / Linux desktop packages, and a Docker self-hosted relay. Login works with an email or a mainland China mobile number, and there's a China mirror at nanomuse.cn if GitHub is slow. The README is available in 10 languages.

It wasn't written from scratch: the Android side is based on 1.13 of OpenMinis (5,783 stars), an open-source phone agent, and OpenMinis's author also contributed code to it. That's how it shipped 41 versions within two weeks of its Sep 25 launch, from Oct 9 to 1.0.

How the stars grew

Star count from GitHub. +371 in the last 7 days, only half of the first 7 days (+726).

The first wave came from the paper and Chinese-language circles. The paper is titled "nanoMuse: an open-source personal agent that runs on every one of your devices." It breaks down how Meta's Muse was built, based on Meta's public materials and a published prompt from the live Muse, then offers an open-source implementation. The paper topped the Hugging Face Daily Papers chart, and Chinese AI media loved the headline "Zhejiang University open-sources a Muse alternative." English-speaking circles barely picked it up; its Show HN only got 68 points.

The second wave came from Trending. The Swift category chart is probably easier to reach than the overall chart (inference). It hit #1 in Oct 21, then gained 485 stars the next day, while a batch of aggregator and tutorial sites indexed it at the same time (the timing lines up, but causation can't be confirmed directly).

It kept giving people reasons to share it. In 44 days it shipped over 50 versions, and each milestone (1,000 stars, 2,000 stars, Trending #1) went into the "News" section at the top of the README. All packages combined have been downloaded about 35k times, and 1.0.5 one version alone 15k times. In Nov 2 @shao__meng shared the paper again (44k views), and over the next two days it gained more than 70 stars each day.

Why this one and not the dozens of other "open-source Muse" projects

After Muse launched, a wave of "Muse alternatives" appeared on GitHub: Comma, xiryn, noodle, MuseLite, pi-muse, and more. Most have tens to four or five hundred stars. Only two passed a few thousand: CopilotKit's openmuse (6,264 stars, out one week after Muse launched) and nanoMuse.

nanoMuse has four things the rest of that group doesn't:

1. A three-column comparison table. The first screen of the README shows Meta Muse, OpenMinis, and nanoMuse side by side, making its position obvious at a glance: Muse runs on Meta's cloud, uses only Meta's models, and can't touch your devices; nanoMuse does the opposite.
2. Academic and institutional backing. A paper with Zhejiang University affiliation, the Hugging Face chart, Xinhua News coverage. An individual developer's "open-source Muse" doesn't get this amplifier. Xinhua's report also says the project was promoted in cooperation with Hangzhou Shuzhi Kunpeng Technology Co., Ltd. and a Zhejiang University lab. What role the company actually plays, and who pays for the free credits, is unclear.
3. Easy to try on sight. Most "open-source Muse" projects only offer a docker-compose file. This one ships installers for six platforms, a browser demo, free credits, and supports mainland China phone-number login.
4. The only one with a phone version. openmuse leans toward web and server-side and doesn't operate phone screens. OpenMinis can operate phones but only manages one device, and it doesn't carry the Muse name.

Still, few people in the discussion talk about how well the features actually work. The stars mostly came from the "Zhejiang University open-source Muse alternative" story. The cross-device design and approval gates are what make that story hold up.

Who it's competing with

Star counts from GitHub, 11-06.

Other projects are already plugging into it: in Nov, 6 open-sourced Amika, which supports connecting to nanoMuse.

Where it falls short

• Almost blank in English-speaking circles. READMEs in 10 languages, but English coverage is only a Show HN with 68 points and scattered reposts, with no major English media coverage.
• Still some friction to install. The macOS package isn't notarized, so the first launch requires right-click → Open; on Windows you have to click "Run anyway"; iPhone only gets a TestFlight beta.
• Hard to tell how many people actually use it. 35k downloads only set an upper bound; there are 120 open issues.

If it were you

Verdict: C. "Open-source Muse" is already taken by openmuse and nanoMuse. Muse's own buzz and star growth are both fading, so building another one now won't get much attention. nanoMuse's paper and institutional backing aren't available to an individual developer. What's worth taking is the playbook, especially the step it adds: when a big-company hit product isn't available in a given region, ship an open-source version that people there can install and use right away.

Who it suits: developers who can assemble an installable app from a mature open-source project within two or three weeks; people who know a market that big-company products keep out (Chinese, Arabic, Southeast Asian); better yet, people in a school or lab who can easily write up a technical report.

Who it doesn't suit: anyone trying to squeeze into the "Muse alternative" space now; people who can only build server-side and can't ship installers; anyone hoping to make money from this.

How to do it: ① Watch big companies' new agent launches (Meta, OpenAI, Google, xAI) and note which countries they can't register in, which platforms they haven't shipped to, and where they've had privacy incidents. ② Fork a mature open-source upstream and deliver the hardest part first (nanoMuse chose phone screen operation). Don't write from scratch. ③ On the first screen of the README, put a three-column comparison of the original product, the upstream, and your version; put installation on the second screen, with a direct download link for each platform, plus a browser demo. ④ Lower the barrier as far as possible: free trial credits, local phone-number login, domestic mirrors.

How to get users: write a technical report that breaks down the original product and post it to arXiv, then go for the Hugging Face Daily Papers chart; in the target market's AI media circles (for Chinese, X Chinese accounts and Zhihu), push a "so-and-so open-source alternative" narrative; pick a language category with less competition and go for Trending; write every milestone into the README's News section, since every release is another reason to repost.

When to move: within two to four weeks after the original product launches. nanoMuse came out ten days after openmuse, and only carved out its own space through the phone version and the Chinese-language market.

Signs you've misread the situation: ① In months when Muse has no incidents, nanoMuse still steadily gains a hundred or two hundred stars a day, which would mean the product itself is pulling people in, not just riding the name. ② After the next big-company agent launches, regional alternatives generally fail to reach a thousand stars, which would mean this playbook has been overused. ③ An individual developer ships a regional open-source alternative of similar scale without a paper or institution, which would mean backing matters less and this kind of case is easier to follow.