# 40 minutes, 100 forks: how a Windows tool that hands Meta Muse to AI agents gained 3 stars in 900 days · GitHub open-source project

> Source: https://stringers.io/en/github/gh_andrew-babo__muse4ai.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> When Meta launched Muse in September and granted free video quota, it spawned a whole category of "turn Muse into an API" projects. But the stars mainly came from a batch of newly registered empty accounts that followed, starred, and forked together, and almost no one discussed it off-platform.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

andrew-babo/muse4ai · andrew-babo

# 40 minutes, 100 forks: how a Windows tool that hands Meta Muse to AI agents gained 3 stars in 900 days

**What it is** A Windows desktop app that uses your own Meta Muse account to generate images and videos, then exposes those capabilities to AI agents to call.

**Why it took off** Meta launched Muse in September with free video credits, which spawned a whole track of 'turn Muse into an API' projects. But its stars came mainly from a batch of freshly registered empty accounts that starred, forked and followed it, with almost no discussion outside GitHub.

**C** **How you can approach it** Don't copy a Muse gateway. Learn its template of 'free generation capabilities + MCP + one integration page per agent', then build the macOS version or video-production workflow it lacks, or wait for the next big company to open free credits and package it in the first week.

C grade: the approach is worth learning, but the Muse track already has a leader with 1,000 stars and a dozen followers, and its lifeblood is in Meta's free-credit hands; this repo's stars are inflated and can't be used as evidence of demand. Few people know about it yet

Data (updated daily) · repo created 2026-11-03 · 351 stars

GitHub · stars

Data 2026-11-05 → 2026-11-08 (4 days, 4 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

Muse4AI is a Windows desktop app built by Vietnamese developer Andrew Babo. You log in with your own Meta Muse account locally, and it packages Muse's image and video generation as MCP and local interfaces so agents like Claude Code, Cursor and Codex can call them directly. The topic catches a real trend and the product is quite complete. But it went from zero to 900 stars in 3 days, and most of that looks bought. If you want to build a similar project to earn stars, you first need to separate these two things.

### Morning of Nov 4, 40 minutes in

The repo was created on the evening of Nov 3, 2026 (UTC). Starting the next day, 06:21, forks began appearing one after another, every few dozen seconds, and in about 40 minutes it had collected the first 100.

Looking at the accounts behind these forks, they look much the same: most registered after mid-October, have 0 followers and 0 to 2 public repos, and carry the suffixes GitHub auto-suggests at signup: -cmyk, -sketch, -lgtm, -a11y, -gif. The names are mostly South Asian, Malagasy and Nigerian. One account registered in Nov 3 has starred a string of dubious repos including CapCut, 'uncensored AI API' and Pika, the typical profile of crowdsourced star-farming accounts.

The author's own followers look the same. andrew-babo registered in March 2025, has only this one public repo, but has 794 followers, nearly as many as its stars. Estimating signup dates from GitHub user IDs, 57% registered after mid-October; a spot check of 8 accounts on the latest page were all new accounts from Nov 1 to 2.

Followers, stars and forks all rise together, and forks make up 21% of stars. This is the standard shape of 'GitHub mutual-help tasks' in Telegram task groups and crowdsourcing platforms (an inference: stargazer details couldn't be checked one by one, so the conclusion comes from the follower and fork lists).

### Metric

| Timeline | stars | fork | Source |
|---|---|---|---|
| 11-03 | Repo created |  | GitHub |
| 11-05 | 468 | 135 | GitHub daily numbers |
| 11-06 | 799 | 171 | GitHub daily numbers, +331 in a single day |
| 11-06 evening | 928 | 201 | GitHub live check |

Almost no noise outside GitHub. Searching for "Muse4AI" with Jina, Exa and Brave returns only the author's own site and GitHub; on X there are 30 posts in the past 0 days; searching TikTok returns only generic Muse videos, none mentioning it. A repo that really spread organically wouldn't go undiscussed everywhere.

Compare the leader in the same track: czg86389-hub/muse2api was created in Sep 25, got reposts from big Chinese accounts on X, made it onto Trendshift, and took 42 days to reach 1,022 stars, gaining 10 to 55 a day. Muse4AI matched that in 3 days with no visible external source.

### What it is, and how well it's doing

First line of the homepage: "Tell your AI assistant what you need in plain words. It drives Muse for you". Say what you want in plain language to your AI assistant, and it operates Muse for you, then saves the finished files back to your computer.

Star count alone doesn't mean the product is an empty shell. The engineering effort is actually substantial:

- The first commit is "2.0.0-beta.31", meaning a finished build that was iterated privately over 30 several versions and then open-sourced in one go; the README says there are 775 tests.

- Local REST and MCP, plus OAuth dynamic registration and 6-bit code authorization; it has an asset library, a task queue, multiple accounts, and video can be extended to 60 seconds.

- The site has a dedicated integration page for each of 18 agents (Claude Code, Cursor, Codex, Gemini CLI, n8n, Manus…), plus llms.txt and a Skill, in both English and Vietnamese.

- The use-case pages are all about selling products: TikTok Shop affiliate videos, Facebook ads for small shops, UGC-style ads, short dramas. The target user is small sellers in Southeast Asia doing e-commerce content.

The author calls it a "non-commercial portfolio project".

### Why the topic itself is real

On 2026 Sep 8, Meta launched Muse, a personal agent with Muse Image and Muse Video. Free accounts get roughly 30 to 50 videos or a few hundred images per week (tested by useapi.net). TikTok immediately filled with tutorials like "free unlimited tokens" and "Unlimited FREE AI Video with Muse", with tens of thousands of plays.

"Turn Meta's free video-generation quota into an API you can call" quickly became a whole track:

| Project | Repo created | Stars | Approach |
|---|---|---|---|
| czg86389-hub/muse2api | 09-25 | 1,022 | Chinese-language, self-hosted OpenAI-compatible gateway, with account-pool rotation |
| www222fff/muse2api | 09-27 | About 340 | Developer-facing, pluggable drivers |
| d4ncboz/museai, pooops15/muse-bridge, imkofty/Hermuse | 9–October | 36–168 | Plugs Muse into other agent frameworks as a text model |
| Muse-Canvas, museflow, muse-studio | 10–November | Single digits to 93 | A canvas and node UI for people |
| bitgacon/muse2api | 10-23 | 18 | Vietnamese-language Windows portable build; positioning closest to Muse4AI |
| lichao199208/muse2api | 11-05 | 32 | Pure protocol reverse proxy, with a registration bot for mass-creating accounts |
| useapi.net Muse API | Around October | Commercial service | Hosted, paid |

The spot Muse4AI occupies really is empty: the others are "developers who can deploy it set up a gateway themselves", while it is "a Windows installer that works right after install, plus native MCP". The difference is real; it's just that its stars didn't come from that difference.

### Its weak spots

- **Stars can't be used as a demand signal.** Bought stars don't mean anyone uses it. Real download counts aren't visible from outside (there are no GitHub releases, and the website doesn't publish numbers).

- **The whole track rests on Meta's free quota.** It goes through web accounts, not official APIs. If Meta tightens the quota, cracks down on reverse-engineered calls, or opens video generation through an official API, the core value of these projects disappears.

- **Windows only.** Developers on Mac using Claude Code or Cursor can't use it.

- **Starring on GitHub carries risk.** If detected, the repo could be demoted or its stars purged (inference).

### If it were you

**One-line verdict:** packaging a platform's free generation capability as an MCP that agents can call is worth learning from, but the specific Muse topic already has more than a dozen projects competing for it, and the window is held by Meta. Building another copy of Muse4AI will struggle to get real stars.

**Who it's for:** developers who can quickly build cross-platform MCP tools and keep an eye on every big lab's new model releases. When the next major company opens free generation quota, you need to have it packaged within the first week.

**Who it's not for:** anyone who just wants to copy a Muse gateway; the category already has a leader with 1,000 stars and a crowd of followers. Anyone hoping to start by buying stars will get neither users nor reputation from them.

**How to do it:**

1. If you still want to build on Muse: do what Muse4AI hasn't done. A macOS version, a one-click MCP extension for Claude Desktop, or chaining "image generation, video generation, and cutting into a finished piece" into an agent workflow, instead of yet another gateway.

2. The better move is to keep this template ready for the next platform event: in the week a big company releases a new generation capability with a free quota, ship a "one-command-install MCP + an integration page for each agent + llms.txt".

3. The top of the README should immediately show a screen recording: an agent call, and within seconds the video lands in a folder.

**How to get users:** follow the path of the leader muse2api, not Muse4AI. Find Chinese- and English-language big accounts on X that post AI freebie tutorials and DM them the recording; post to the matching subreddits and MCP directories; Vietnamese and Indonesian e-commerce seller groups and Facebook groups are where real users are.

**How long you have:** this Muse wave will likely be rewritten by Meta's rules or an official API within two or three months (inference). If you're doing it, ship this month. The "next platform event" has no deadline, so the template can be set up now.

**What would show this read is wrong:** if Muse4AI draws real discussion, tutorials, and issues off-platform without inflated metrics, that means the "no-deploy Windows version" gap is bigger than it looks and a better version is worth building seriously. Conversely, if Meta tightens the free quota or blocks web calls before year-end, this whole track stops in its tracks.

[GitHub repo ↗](https://github.com/andrew-babo/muse4ai)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| czg86389-hub/muse2api | Open source | 2026-09-25 | About 1000 stars, listed on Trendshift, shared by a big Chinese-language X account | Self-hosted OpenAI-compatible gateway + account pool; Muse4AI is a desktop app that needs no deployment, with MCP/OAuth |
| www222fff/muse2api | Open source | 2026-09-27 | About 341 stars | Developer-facing gateway, pluggable drivers |
| bitgacon/muse2api(Muse2API Studio) | Open source | 2026-10-23 | 18 star | Also a Vietnamese Windows portable build, but leans toward Telegram and manual use; no MCP/OAuth |
| lichao199208/muse2api | Open source | 2026-11-05 | 32 star | Pure protocol reverse proxy, with a registration bot for mass-creating accounts |
| d4ncboz/museai, pooops15/muse-bridge, imkofty/Hermuse | Open source | 2026-09~10 | 36–168 star | Plugs Muse in as a text model for 9Router and Hermes; doesn't focus on image or video generation |
| Muse-Canvas / museflow / muse-studio | Open source | 2026-10~11 | Single digits up to 93 stars | A canvas and node creation UI for people, not for agents |
| useapi.net Muse API | Web | Around 2026-10 | Commercial hosted API; scale unverified | Hosted, paid; Muse4AI is free and runs locally |
| Meta's official Model API (Muse Spark) | Big-company features | 2026 | Official | Official paid API; whether it includes video generation is unverified. Muse4AI goes through web-account free quota |

Where users come from

- Inference **Paid crowd-sourced follow/star/fork** : of 794 followers, 57% were registered after about 2026-10 (estimated from user IDs); the newest followers are all empty accounts registered in 11-01–02; the first 100 forks appeared in a continuous burst within about 40 minutes starting at 11-04 06:21, and 58% are new accounts registered after mid-October (suffixes like -cmyk, -sketch, -lgtm; names from South Asia, Madagascar, Nigeria); follower count and star count are nearly identical

- **Verified** **Platform event (Meta Muse free quota)** : Meta launched Muse and Muse Image/Video in 2026-09-08; free accounts get roughly 30–50 videos a week (measured by useapi.net); "free access" tutorials on TikTok have tens of thousands of views

- **Verified** **Own-site SEO / agent integration pages** : muse4ai.andrewbabo.com has a separate integration page for each of 18 agents, plus llms.txt, use-case pages, and a Vietnamese version; so far, search has indexed only its own site

- Inference **Spread on X / Reddit / HN / TikTok and other channels** : x-radar, TikHub, Exa, and Brave found no content mentioning Muse4AI

## Research report on andrew-babo/muse4ai (Muse4AI)

Research date 2026-11-06. Channel: GitHub.

### One line

Muse4AI is a Windows desktop app (Electron + TypeScript, MIT). Users log into their own Meta Muse (muse.ai) web account on their machine, and the app wraps Muse's image generation, video generation, chat, and video extension/editing into a local REST API (127.0.0.1:47821), MCP, and OAuth interfaces for agents like Claude Code, Cursor, Codex, and n8n to call. The author, Andrew Babo, is a Vietnamese developer; the site and docs are bilingual in English and Vietnamese, and he calls it a "non-commercial portfolio project".

**结论:** 它有真实的时机红利。Meta 2026-09-08 发布 Muse,带免费的出图、出视频额度,随后冒出一大批“把 Muse 逆向成 API”的项目(muse2api 一系)。Muse4AI 抢的位置是“给 agent 用的、不用部署的 Windows 版 + MCP/OAuth”。但 3 天拿到 799–926 star 这个速度,大部分很可能来自付费的「关注+star+fork」众包任务,不是自然传播:关注者有一半以上是 10 月中以后注册的空号,fork 账号集中在孟加拉、巴基斯坦、马达加斯加、尼日利亚一带的新号,站外几乎零讨论。这个库的 star 不能当作需求信号。

### Data

- Data card (GitHub repo): 11-05 468 stars / 135 forks, 11-06 799 stars / 171 forks, +331 in a single day, fork/star ratio 21%.

- GitHub API check (11-06 later): 926 stars, 201 forks, 16 watchers.

- Repo created 2026-11-03 17:45 UTC. The first commit is "Release Muse4AI 2.0.0-beta.31 source for the open-source community", meaning a finished product that had already gone through over 30 betas privately, then open-sourced in one go. As of 11-06 there are 21 commits in total (beta.32, beta.33, an audit report, Dependabot). The README says there are 775 tests.

- The author account was registered 2025-03-19 and has only this one public repo, yet has 794 followers, nearly the same as the star count.

### Growth tactics: the evidence points to inflated metrics

1. **Follower makeup.** We pulled all 794 followers of andrew-babo and estimated registration dates from GitHub user IDs (for example, the account hinnakha11-cmyk with ID 341,804,059 was registered in 2026-10-13): 455 (57%) were registered after about mid-October, and 540 (68%) after about August. A spot check of the latest page covered 8 accounts, all registered between 2026-11-01 and 02, with 0 followers each, 0–2 public repos each, and usernames carrying GitHub's auto-suggested suffixes (-beep, -rgb, -boop, -byte, -maker).

2. **fork accounts.** The first 100 forks landed every few dozen seconds starting 2026-11-04 06:21, all within about 40 minutes. 58% of them are accounts registered after about October, with names like -cmyk, -sketch, -lgtm, -a11y, -gif, -netizen. Most have South Asian, Malagasy (Koloina04, manalazyfitiavana, kennyandriatsilavo-lang), or Nigerian names. For example, toria9673-lgtm signed up 11-01, and melovefast4336 signed up 11-03. The repos melovefast4336 starred also include KiloJellyfish99/CapCut, UncensoredLabs/uncensored-ai-api, Pika-Labs/Pika-H3-Flex, and a string of other suspicious repos. This is a textbook crowd-sourced star-farming pattern.

3. **Follow, star, and fork all rising together.** This is the typical shape of microworkers and Telegram task groups offering "GitHub help" or paid tasks. A fork/star ratio of 21% is high, which fits a task requirement to "fork it while you're there."

4. **Almost no organic spread off-platform.** Searching "Muse4AI" on Exa, Brave, and Jina only returns the author's own site and GitHub. x-radar shows 30 posts mentioning "muse4ai" in the last 0 days. Searching muse4ai on TikTok (via TikHub) returns generic Muse videos, and none of them mention it.

5. **Comparison.** The hottest repo in the category, czg86389-hub/muse2api (Chinese-language, created 9-25), went from 42 days to about 1000 stars, gaining 10–55 per day (github repo_day, 10-23~11-06). www222fff/muse2api reached about 340 stars in 40 days. Muse4AI caught up with the leader in 3 days without a single external traffic source, which is hard to explain with organic demand.

Note: the GitHub stargazers endpoint returned 404 for this machine's token (and 404 for czg86389-hub/muse2api too), so it's an API issue unrelated to this repo, and I couldn't check stargazers individually. The conclusions above come from the followers and fork lists, so they are strong inferences, not individually verified.

### Why people want it (the real-demand half)

- **Platform event.** Meta 2026-09-08 launched Muse, a personal agent ([Wikipedia](https://en.wikipedia.org/wiki/Muse_(AI_agent), [Meta press release](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/)), with Muse Image and Muse Video built in ([Meta AI blog](https://ai.meta.com/blog/introducing-muse-image-muse-video-msl/)). Free accounts get about 30–50 videos a week or a few hundred images ([useapi.net tests](https://useapi.net/docs/articles/muse-images-videos)). TikTok has tutorials like "free billion tokens" and "Unlimited FREE AI Video with Muse" with tens of thousands of views (TikHub search). "Getting free Meta video credits" is the shared driver behind this batch of products.

- **Where it sits.** The muse2api family is self-hosted OpenAI-compatible gateways (Python, Docker, account pools) aimed at developers who can deploy them. Muse4AI ships a Windows installer that works out of the box, adds MCP (compatible with OAuth dynamic registration and 6-digit device codes), a local asset library, queuing, multiple accounts, and video extension up to 60 seconds, and writes a setup page and a Skill for each of 18 agents. Its target users are small sellers in Vietnam and Southeast Asia making TikTok Shop affiliate videos and Facebook ads (the site's use cases are all affiliate videos, UGC ads, and short dramas).

### Competitors

| Name | Form | Launched | Scale | How Muse4AI differs |
|---|---|---|---|---|
| czg86389-hub/muse2api | Open-source Python gateway | 2026-09-25 | About 1,000 stars (github repo 11-06), listed on Trendshift, reposted on X by a large Chinese account (@wlzh) | Self-hosted OpenAI-compatible API with account-pool rotation; Muse4AI is a desktop app + MCP, no deployment needed |
| www222fff/muse2api | Open-source gateway | 2026-09-27 | About 341 stars | Pluggable drivers (browser CDP / http), aimed at developers |
| lichao199208/muse2api | Open-source pure-protocol reverse proxy + registration bot | 2026-11-05 | 32 star | Ships batch-registration of alt accounts, distributed through QQ groups |
| bitgacon/muse2api | Vietnamese-language Windows Studio | 2026-10-23 | 18 star | Also Vietnamese, Windows portable build, Telegram; the closest positioning match |
| d4ncboz/museai, pooops15/muse-bridge, imkofty/Hermuse | Open-source gateway / bridge | 9–October | 36–168 star | Connects to 9Router and Hermes, using Muse as a text model |
| huwanbin/Muse-Canvas, bynne2602/museflow, binance9/muse-studio | Canvas, node-based workflows, one-click Vietnamese video generation | 10–November | Single digits up to 93 stars | A creator UI for people, not for agents |
| useapi.net Muse API | Hosted paid API | Around 2026-10 | Commercial service (scale unverified) | Hosted, no computer needed; Muse4AI is free and runs locally |
| Meta's official Model API (Muse Spark) | Big-company features | 2026 | Official | Official paid API; whether it offers video generation is unverified (TriTue2011/muse2api is a gateway built on it) |

### Visibility

Almost nobody knows about it. There is zero off-platform discussion, no media coverage, nothing on X, and no one is copying it. The whole muse2api category circulates in small Chinese, Vietnamese, and Indonesian circles. The leader, czg86389-hub/muse2api, is listed on Trendshift and has been reposted by a large X account.

### Differentiation

The product differentiation is real: "no deployment, one-click Windows install, native MCP/OAuth, aimed at agents and non-developer small sellers, Vietnamese language." The engineering effort is also substantial: 33 betas, 775 test suites, and a full docs site. But that differentiation is not why the stars climbed. The climb came mainly from inflated numbers. Whether the differentiation can win real users is not yet supported by any evidence.

### What others can borrow

- When a platform offers free quota (Muse video, images), the fastest business is to wrap it as an MCP or desktop installer that agents can call, for people who can't deploy anything.

- The playbook of "one access page per agent + llms.txt + Skill" for SEO / agent-SEO.

- Conversely, this is also a case study in star farming: follows, stars, and forks rising together, with fork accounts that are newly registered empty shells. When you see a "new account, single repo, thousands of stars in 3 days, zero off-platform buzz" pattern, check the followers and fork lists first.

### Uncertain

- The real download count of the installer is unknown (the site doesn't publish it, and GitHub has no releases).

- Whether the inflated numbers were bought by the author or were gifted by someone else cannot be confirmed.

- Whether Meta will block this kind of reverse-engineered usage (the site itself states it "does not circumvent restrictions"). Whether the official API has video features is unverified.

- Stargazer details could not be retrieved (the endpoint returned 404).

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.