Turning Zuckerberg's 30B tokens into an API: muse2api got 42 stars in 1000 days, thanks to going live two days earlier

Meta launched Muse in September. New accounts got 1B tokens, and invite codes could stack that to about 30B. But the credits could only be used in Muse's own app and website. A creator with just one repo and 18 followers created muse25api in Sep 2. It wraps the Muse web interface in an OpenAI-compatible API and lets you pool a bunch of accounts and rotate through them. By Nov 6 it had 1,026 stars and 287 forks. Below: what it is, where the stars came from, why this one took off when a dozen similar projects didn't, and whether you could still make a project that gets stars the same way.

"Holy crap! It worked!!"

In Sep 27, a Chinese AI blogger on X with 15K followers, Kin (@KinGao476942), posted: "Holy crap! It worked!! I've successfully hooked Muse's 30B token credits into my Grok bot! Writing articles, generating code, making images, making videos, all now run on Muse's credits from inside Grok bot... Zuck, I bet you never imagined we'd use your credits like this!"

Below is a three-step tutorial, and step one is "send the muse2api project link to your Grok bot." The post got 75K impressions, 512 likes, and 723 bookmarks (on X, 11-06). More bookmarks than likes means people saved it to follow later. The next day, a repo that was only three days old gained 68 stars in one day (trendshift).

What it is

It's a Python program that opens the Muse website in headless Chrome, types prompts, generates images and videos for you, and returns the results in OpenAI's API format. So anything that can connect to OpenAI, like Grok bot, Codex, various infinite canvas tools, or your own scripts, can use the credits Muse hands out.

It differs from other "Muse-to-API" projects in two ways:

• It covers chat, images, and video. Text-to-image, image-to-image, text-to-video, and first-frame image-to-video are all supported. Meta's official API only sells text models (Muse Spark at $1.25 / $4.25 per million tokens), and the free credits in the app aren't usable there either. If you want reliable access to Muse video, the web route is basically the only option.
• Built for people with lots of accounts. Import as many accounts as you like and rotate through them. When one account's credits run out, it switches to the next. Sessions renew automatically every 48 hours, and each account can have its own proxy. There's a Chrome extension that imports whichever account you're logged into with one click, plus a dark-mode admin panel. When the author publishes a new release on GitHub, every machine running it gets a one-click online upgrade prompt.

On the LINUX DO forum, a user posted in Oct 25, "Made five more MuseAIs." The first reply was just "Reverse-proxied 8 accounts, and we still can't use up the weekly credits," with two panel screenshots attached. The thread grew to 105 replies, and someone in it directly recommended "try this project, czg86389-hub/muse2api." That's its user base: not people who want to use one account, but people who've collected a dozen or more.

Where the stars came from

Since the repo was created 9-25, its daily star growth has had two peaks:

Data from trendshift and GitHub daily snapshots. The last 7 days added +239, the first 7 days added +237, and it has held steady at over thirty a day.

The first wave was the coupon crowd. Muse launched in Sep 8 and hit #1 on the US App Store. A whole Chinese ecosystem grew around it quickly: invite-code swap groups on Zhihu, Telegram bots that farm invites, YouTube tutorials titled "Free TG tool to grab 30B tokens" and "residential IPs to avoid bans," and platforms selling ready-made accounts. What this crowd lacked was a layer that took the credits out of the app, and Kin's post made muse2api the default answer.

The second wave was video. In late October, AI video plus infinite canvas tools were trending, and people found Muse's video could stand in for some paid models. On X, @oh_Roach 11-03 wrote, "Using Muse2API, you get a Seedance2.0 substitute for generating video... drop it into your infinite canvas as a model provider... also supports matrix accounts." @dragonvsmm 11-04 was blunter: "Build your own account pool, run infinite canvas for video. It's dozens to a hundred times cheaper than sd2.5, and if they ban the account, just swap it."

Even AI started recommending it. Starting in November, the @grok account on X replied to several Chinese posts saying, "For batch Muse accounts to build a pool, just use muse2api directly (it's on GitHub, I'd recommend the czg86389-hub one, with a web console and automatic keepalive)."

Vietnam too. In the Facebook group stablediffusion.vn, someone explained "making TikTok videos from start to finish with Muse2API + Claude," and Vietnamese developers built several Vietnamese-language forks.

The English-speaking world has barely noticed it. In the last 7 days, the 2 posts on X mentioning muse36api were almost all in Chinese, the most-liked one had 20 likes, and no media outlet covered it.

How many people are actually using it? Forks are 28% of stars, while new repos of similar size usually sit around 11%, which suggests plenty of people pulled it down and deployed it. Another hint is the attachment download count for the Chrome import extension: about 960 downloads total across v1.6.3 through v1.6.9 (GitHub, 11-06). Instances that are actually running and keeping up with upgrades probably number in the hundreds to a thousand (estimate).

A dozen same-named projects, why this one

At least 15 open-source implementations of "Muse web-to-API" exist on GitHub.

There's also the paid useapi.net ($15 / month hosted, 3 accounts per service).

Three things created the gap. One: early. 9-25 created the repo, two days before the same-named projects, and it caught Kin's post. Two: it actually works. Video was made to work first, and account pooling, renewal, proxies, and the import extension were all built together. The frequent issues in the issue tracker, like ban risk, import errors, and video misjudgments, are being fixed one by one. Three: fast iteration. 41 days and 61 commits, 9 releases (v1.6.1 to v1.6.9,11-06 shipped two versions in one day). Besides the author, about 10 other people submitted PRs, adding Tampermonkey-based account import, per-account proxies, an English interface, tool calling, and multi-reference images. The README pins a LINUX DO community badge, and the community has embraced it.

Others are already following

Forks and rebrands are already piling up: a Vietnamese Studio version (bitgacon/muse2api, 18 stars 23 forks), a Vietnamese-Chinese bilingual version with a one-click Windows package, a React panel version, muse2api with infinite canvas (huwanbin/Muse-Canvas), a command-line version, and a small cookie-import tool. Their stars range from single digits to a dozen or so.

One step further is lichao199208/muse2api, which appeared only in 11-05: no browser, a built-in registrar, aimed straight at mass-producing accounts. That suggests the next stage of this path belongs to account resellers, not open-source projects.

If it were you

C: worth learning from, not worth chasing. Muse's spot is already held by this project, with a dozen-plus similar ones, and the later ones all sit under a few dozen stars. The whole thing is also built on a Meta loophole. Meta is already banning accounts and adding face verification, and @grok has started refusing to help people deploy "credit-draining" bots (11-04 someone posted a screenshot). The part worth taking is the playbook: when the next big company hands out credits, be the first to build everything out.

Who it suits: developers who already have working code for "web-to-API + account pool + import extension" and can adapt it to a new platform within a day or two. Ideally you have some presence on LINUX DO or the Chinese X community, or know one or two bloggers who write tutorials.

Who it doesn't suit: anyone who wants to build another reverse proxy on Muse, since that spot is gone. Anyone hoping to overtake it with more features and a cleaner architecture. www222fff is an example: two days later, with a better architecture, and it has only a third of the stars. And anyone worried about the project being taken down or called out by the platform.

How to do it: ① Watch for new products where big companies hand out huge free credits but only open an app or website. You need to be working within a week of launch, and being two days early is the whole difference. ② Build account pooling from day one, not single-account support. People farming coupons hold a dozen-plus accounts, and a single-account reverse proxy doesn't reach them. ③ Make account import one-click (a browser extension or Tampermonkey script). This is the highest-barrier step. ④ Get video and image generation working first. A modality others can't offer is the easiest to spread as a "substitute." ⑤ Put a one-click upgrade in the panel. When upstream changes, ship a new release, and users won't drift to other forks.

How to get users: Find a Chinese blogger who writes tutorials and give them a concrete use case they can follow, like "plug the free credits into a Grok bot" or "a Seedance substitute." That spreads much better than "OpenAI-compatible." Then post on LINUX DO and display the community badge. In communities that are active about coupon hunting, like Vietnam's, forks will grow on their own.

When the window is: For this kind of project, the window opens one to two weeks after a new platform launches, and the attention you can capture usually lasts a month or two. For Muse, that window has already closed. The next one depends on when the next company hands out credits.

Signs you've read it wrong: ① Even after Meta patches the loophole or revokes the free credits, muse2api's stars keep rising steadily. That would mean its value is more than coupon farming, and people really use it as a video API. That's worth taking seriously as a "web video model to API" product. ② A pure-protocol version with a registrar, like lichao199208, overtakes it within a month. That would mean what's valuable on this path is the ability to produce accounts, not engineering polish. ③ An English version gets several hundred stars on Hacker News or Reddit. That would mean the English-speaking gap is real, and the value of translating and following up has been underestimated.