2,000 stars, 200 downloads: how a Douyin livestream review workbench got its stars
What it is An open-source Windows desktop app that helps Douyin livestream teams find hosts, turn livestream recordings into transcripts, run reviews by aligning scripts with minute-level traffic, and write short-video scripts.
Why it blew up It didn't, really: the stars were most likely bought. The accounts forking it are batch-registered alts, and the same batch was also boosting a group of Chinese AI Agent repos. When GitHub purged fake accounts on 10-23, it lost its stars along with those repos. Real downloads: only 210.
DHow you could follow this Don't chase stars by copying the workbench. If you really want this niche, build one small, solid open-source library that aligns livestream recordings with platform traffic tables to produce script reviews, supports TikTok Shop and WeChat Channels, and promote a full breakdown of one public livestream in operator communities.
- Data (updated daily) · Repo created 2026-08-22 · 2,037 stars · 7 days +268
GitHub · stars Data 2026-10-09 → 2026-11-08 (31 days, 31 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
On Aug 22, 2026, an account called HanyuanWang created a GitHub repo named LiveStream Agent Studio. Forty-some minutes later it posted a 368MB Windows installer. The next day, among the first accounts to fork it was one called fans3rvice, which reads as "fan service." 76 days later it had 2,036 stars, 217 forks, and the installer had only been downloaded 210 times. This piece covers what it does, where the stars came from, and how much window is left in this opportunity. Stringers
What it is
A Windows desktop app for Douyin livestream sales teams, open source under the MIT license. It chains four Agents:
- Host discovery: reads the Chanmama rankings and host pages visible to your account, and compiles candidate hosts and creator breakdowns.
- Livestream breakdown: extracts audio from a full livestream recording and sends it to Alibaba Cloud Bailian for transcription, outputting a transcript Excel with second-level timestamps.
- Livestream review: aligns the transcript with the minute-level traffic table exported from Jinritemai (Douyin's merchant backend) for the same stream, showing which lines of script line up with changes in entries, exits, dwell time and product clicks, and outputs Excel and Word reports.
- Short-video scriptwriter: paste a reference short-video link, extract the transcript, and rewrite it as an original spoken script and shot list.
You have to configure your own Alibaba Cloud key and storage. The README is bilingual, the version badge says v0.2.0-beta, but the Release is v1.0.0.
A 'delivery package' left in the repo
Its docs folder contains a 'GitHub Project Introduction Copy' whose first line reads: 'This document is for creating and publishing the LiveAgent Studio repo for the first time. Content in angle brackets needs to be replaced, e.g. <your GitHub username>.' Further down: 'Open the following in the delivery package: 01_GitHub上传内容/LiveAgent-Studio/'. Another doc, 'GitHub Publishing Manual', tells you to set 'Repository name' to liveagent-studio and 'it's recommended to choose Private first'.
In other words, this is a package delivering code and publishing copy together, uploaded as-is, manuals included (inference: AI-generated or outsourced delivery). 12 commits, the last on Aug 24 to edit the README, then no changes for over two months. The 10 PRs dependabot opened on the day the repo was created were never merged. In early September, 3 people opened issues and PRs fixing a hardcoded OSS address and a React dependency; the author never replied.
Where the stars came from: a batch of accounts registered the same day
Look at who's forking it. The recent forkers, semmamhaygood, sammoek1, and ebrycesqlopez, all signed up in Apr 16 of 2025, and xsmith-marinah in Apr 15. Each account has 0 to 1 followers and about twenty or thirty repos, all of them forks (GitHub API, 11-06). fans3rvice also has 0 followers and 33 repos, all forks.
What's more telling is what they forked. These accounts' fork lists overlap heavily, repeatedly hitting the same group of repos: hestjs, ai-shortVideo-pipeline, sell-sessions-skill, growth-lab, Meldwork, open-compute, clowder-ai, plus two 'chatgpt-plus-pro-codex-cn-guide' repos by different authors. These repos look alike: created in 2026, authors with only one or two repos, one to five thousand stars, very few commits. It looks like a customer list for the same star-inflation service (inference).
Several numbers line up with this read:
| Signal | Metric |
|---|---|
| curve | 9-07–09-09 gained +65, +64, +25 per day, then from 9-10 to 9-30 it gained only about 15,10-01 over three weeks, then went back to +20–40,11-03 a day, and from 11-05 to +53–62 a day (daily data from Trendshift). None of the jumps match any spreading event. |
| 10-23 star drop | On that day it fell from 1,520 to 1,477, losing 43 stars. In the same batch of repos, sell-sessions-skill lost 41 that day and open-compute lost 569. Among the GitHub repos we track, 20% lost more than 1.4 stars that day, versus a normal 0.1%–0.4%. It looks like GitHub ran a batch purge of suspicious accounts and this repo was caught in it (inference). |
| Downloads | A desktop app you must download and install to use, with 2,036 stars against 210 downloads (Release ZIP, GitHub API, 11-06) |
| Discussion | In the last ~90 days, searching the repo name on X returned 0 posts; I found no dedicated discussion on HN, Reddit, Xiaohongshu, Douyin, WeChat Official Accounts in the livestream e-commerce scene, or Zhihu. |
The three-week lull in the middle looks most like a pause between a bought package running out and renewal (inference). The fork-to-star ratio is 10.7%, close to similarly aged new repos, because this service inflates forks along with stars, so the common rule of 'stars with no forks' doesn't catch it.
The GitHub API doesn't expose the stargazer list, so I couldn't check the star accounts individually. The judgments above rest on the forking accounts, the shape of the curve, and the synchronized star drop. How many of the roughly 770 stars from the first two weeks came from real people is unclear.
Those exposures were results, not causes
Searching its name turns up plenty: a LinkedIn 'GitHub AI Agent Trends' leaderboard post reading '782 stars (+391 this week)'; AI tool directories like skillsllm.com, theagenticleaderboard.com and agenticmart.org; an Instagram Reel from @github_dev titled 'One App, Four Livestream Agents', with #Mac in the caption (the app only supports Windows); an Arabic-language Facebook page and a YouTube Short, 'Automate Your Entire Livestream Business With AI'.
All of these are automated leaderboards and repost accounts that scrape repos by weekly star growth. They added it because the stars were already rising, not the other way around. I found no promotion by the author themselves.
Is there anything to the product idea?
Somewhat. 'Matching transcripts against minute-level traffic' is a real need in livestream review. Chanmama's own tutorials describe it, and the DIY method is basically a transcript with timestamps. Among open-source projects, it was the only one I found that combines host finding, transcription, script-by-traffic alignment and script writing into one desktop workbench.
But people are already charging for this need:
| Who | What it is | How it compares |
|---|---|---|
| Chanmama | Douyin e-commerce data SaaS, with AI review and highlight scripts | It's itself a data source; this repo scrapes its pages |
| Jinritemai / Douyin e-commerce compass | Douyin's official backend | Replays, per-minute traffic, and highlight moments are all there; this repo's review input is the table exported from Jinritemai |
| Aifuping, Feigua, Douchacha, Huitun, Changuanjia | Paid review SaaS, rose through 2020–2024 | Aimed at brands and outsourced operators; screen recording plus people/goods/venue review |
| Feishu Multidimensional Table 'AI Review Master' | Big-tech plugin, 2025 | Upload screenshots and get a review, with a lower barrier to entry |
| ihmily/DouyinLiveRecorder | Open source, 11,180 stars | Only records livestreams, no analysis |
| NanmiCoder/MediaCrawler | Open source, about 70k stars | Data-collection infrastructure for platforms like Douyin |
On the open-source side, the projects that really did get lots of stars are low-level tools for recording and collection. People who go as far as review need data sources and stability; they pay SaaS vendors, and they won't go to GitHub to download a two-month-stale installer.
If it were you
D. The repo's stars say nothing about demand. If you copy it, you won't get the same stars unless you buy them too. Bought stars get purged when GitHub cleans up accounts, and insiders can spot them by sampling a few forkers, which hurts your reputation.
Who it suits: developers already running Douyin, TikTok Shop or WeChat Channels livestream operations with a real review workflow, who want a tool their own team will use. The stars are a side effect.
Who it doesn't suit: people hoping to rack up stars fast with 'Chinese vertical Agent + bilingual README + one-off release'. This project already showed that path; its stars were bought, not earned by the packaging.
How to do it: if you really want this niche, don't build yet another 'four-in-one workbench'. Build its hardest piece: a CLI or Python library that takes a livestream recording and a platform-exported minute-level traffic table, and outputs an alignment table of script segments against traffic changes. Support platform export formats beyond Jinritemai, adding TikTok Shop and WeChat Channels, and let transcription use local Whisper so it isn't tied to Alibaba Cloud. Keep it small, pip-installable, and ship a sample report from a real livestream.
How to get users: post the sample report where livestream operators gather, not the GitHub link: under posts about livestream review on Xiaohongshu, Jike and Zhihu; on the English side, Reddit and Discord for TikTok Shop sellers. Use a full breakdown of one public livestream as the demo. It works better than any README.
When is 'before' now: no window pressure. The need won't disappear, but there's no opening that's currently closing either.
What would mean I got it wrong: (1) the Release downloads climb past a thousand within a month, or real people keep opening issues, which would show a real group of people hunting for an open-source livestream review tool; (2) Chinese livestream-commerce circles (WeChat Official Accounts, Xiaohongshu) have people writing up hands-on experiences with it; (3) a second similar project appears in open source and picks up a few hundred stars in a month without any boosting.
One habit that's easy to use: when a new Chinese repo's stars climb fast and evenly, open its fork list, check the sign-up dates and what those forkers have forked, then look at the Release download count. In a minute you can tell whether people are using it or people are buying it.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one Chanmama (AI review, high-light talk scripts) Web From ¥2019 A leading Douyin e-commerce data SaaS, paid membership It is a data source itself and also has AI review; this project only scrapes Chanmama's pages, then does its own transcription and alignment Oceanengine Buyin / Douyin E-commerce Compass / livestream service platform post-broadcast review Big-company features Douyin official Available to all merchants The official side offers replays, per-minute traffic, and highlight moments; this project takes the official per-minute traffic table as input Ailf / Changuanjia / Feigua / Douchacha / Huitun Web 2020–2024 Commercial paid tools, aimed at brands and outsourced operators Mature screen-recording plus people-goods-field review SaaS; this project is open source, runs locally, and has users bring their own DashScope/OSS keys Feishu Multi-dimensional Table AI Review Master Big-company features 2025 Big-company ecosystem (Jiao Ge Friends' double 11 case studies) Upload a screenshot and get a review out, a lower barrier to entry Tike AI Agent (livestream clip-and-assist Agent) Other 2025–26 Small, listed on the Microsoft Store Leans toward real-time in-stream assistance; this project leans toward post-broadcast review ihmily/DouyinLiveRecorder Open source 2023-07 11,180 star Recording only, no AI analysis NanmiCoder/MediaCrawler, Evil0ctal/Douyin_TikTok_Download_API Open source 2023 / 2021 70K / 21K star Foundational infrastructure for Douyin scraping and downloading, not a review workflow xixihhhh/clipforge, nexscope-ai/eCommerce-Skills Open source 2026 1.2K / 1.3K star Also e-commerce AI, but it does product-pitch short-video generation and cross-border skills, not livestream review - Where users come from
- Inference Paid star/fork account pool: the forkers are 2025-04-15/16 accounts registered in the same batch, 0 followers, and their libraries are all forks (semmamhaygood, ebrycesqlopez, sammoek1, xsmith-marinah). On the first day one forker was named fans3rvice. The same batch also forked hestjs, sell-sessions-skill, growth-lab, Meldwork, open-compute, chatgpt-plus-pro-codex-cn-guide, and other repos from the same group. 10-23 this repo lost −43, and sell-sessions-skill from the same pool lost −41 and open-compute −569 that day. Across the whole repo set, 1.4% of repos lost more than 20 stars in one day, versus only about 0.2% on a normal day. The curve rose at a steady rate, with a three-week stall at 0.
- Verified automated directory sites and star leaderboards: indexed by skillsllm.com, theagenticleaderboard.com, and agenticmart.org; Dr. Wei Liu's LinkedIn post "GitHub AI Agent Trends" lists "782 stars (+391 this week)"; all of these index it because of its star count
- Verified overseas reposting accounts: an Instagram @github_dev Reel (the caption uses #Mac, the wrong platform), the Arabic-language Facebook page slyverex, and a YouTube Short
- Inferred X / media / HN / Reddit / Chinese e-commerce circles: x-radar shows 90 posts in 0 days; searches turned up no related coverage or discussion
HanyuanWang/LiveStream-Agent-Studio: a local Windows AI workbench for Douyin livestream e-commerce
Research date 2026-11-06. Repo github.com/HanyuanWang/LiveStream-Agent-Studio, MIT, no homepage.
Bottom line
它的 star 不是用户、也不是传播带来的。最可能的来源是一个付费刷 star / fork 的账号池:替它 fork 的账号批量注册、没有粉丝,而且同一批账号还在给同一批中文「AI Agent」新库 fork;10-23 GitHub 清理可疑账号那天,它和同一批库一起掉了星;曲线是 76 天「匀速」上涨,而 Release 下载只有 207 次。产品本身(蝉妈妈抓取、云端转写、话术×分钟流量对齐出报告,串成一个工作台)有真实的垂直需求,但没有证据表明这个需求带来了 star。召回理由说它靠中文 AI 榜单、资讯站、LinkedIn 大 V 放大,我们查到的情况相反:那些曝光都是自动聚合站在跟着 star 数收录它,是 star 的结果,不是 star 的来源。
1. Data
- Data card / our GitHub repo (repo_day): built on 2026-08-22; 10-17 had 1,378 stars, 11-06 at 1,984 (11-06 later, the API showed 2,036 stars and 217 forks). gain_7d is 260, and the previous 7 days were 247. It never made trending, but that tells us nothing, since repos younger than about 60 days can't make trending.
- Trendshift daily data (https://trendshift.io/repositories/181558): 09-07 +65, 09-08 +64, 09-09 +25, 09-10 to 09-30 was almost 0 (about 15 over three weeks). From 10-01 on it went back to +20~40,11-03~05 a day, then +53/+56/+62 a day. Over 61 days it gained 1,212 in total. Working backwards, the first two weeks of the database (08-22~09-06) added about 770, which matches the LinkedIn post's "782 stars (+391 this week)".
- Commit history: 12 commits in total, the last push was 2026-08-24, and it was only a README edit. No code changes for over two months after that. 08-22 On the day the repo was created, dependabot automatically opened 10 PRs, and none of them were merged.
- Release v1.0.0 (08-22) has only one 368 MB ZIP, downloaded 207 times, and the sha256 file was downloaded 63 times (GitHub API). Compare that to stars: ten thousand stars should map to about a thousand downloads, which is absurdly low for a desktop app that requires downloading an EXE to use.
- Issues: only 3 real issues / PRs from people (08-31~09-06, all small fixes such as a hardcoded OSS endpoint and React dependency problems), and the author never replied.
- Author HanyuanWang: account registered 2018 years ago, has only this one public repo, 179 followers, and no bio. Co-author JialiangFu has 0 repos and 0 followers.
- The repo's docs/ folder contains "GitHub发布操作手册.md" and "GitHub项目介绍文案.md", which include the lines "open the 01_GitHub上传内容/LiveAgent-Studio/ folder in the delivery package", "set the Repository name to liveagent-studio", and "<your GitHub username>". This looks like an AI-generated or outsourced "code plus launch copy" package that was uploaded as-is. The README's version v0.2.0-beta doesn't match the release's v1.0.0, which supports this.
2. Boosting evidence (the key part)
- Fork account pool: the forkers look like a batch of throwaway accounts. Spot-checking recent forkers: semmamhaygood, ebrycesqlopez, and sammoek1 were all registered in 2025-04-16, and xsmith-marinah in 2025-04-15. Each account has 20~40 repos, all forks, and 0~1 followers (GitHub API). Among the first batch of forkers on the second day after creation (08-23) was an account named fans3rvice ("fan service"), with 0 followers and 33 repos, all forks.
- These accounts fork the same set of repos: the fork lists of semmamhaygood, ebrycesqlopez, sammoek1, xsmith-marinah, buryeo, moukiss and others overlap heavily. Common ones include hestjs, ai-shortVideo-pipeline, Loop-Engineering-for-VLA, QVerisFlow, sell-sessions-skill, growth-lab, Meldwork, open-compute, and clowder-ai, plus two same-theme repos from different authors, "chatgpt-plus-pro-codex-cn-guide" (ChatGPT Plus reselling traffic in China). In our github repo data, this group shares a profile: created in 2026, each author has only 1~2 repos, 1,000~5,000 stars, a fork-to-star ratio of 3%~10%, and very few commits. This is a customer base for a "Chinese AI Agent new-repo star boosting" service.
- 10-23 synchronized star loss: our repo data shows that on 10-23 this repo lost −43 (1,520→1,477). On the same account pool, sell-sessions-skill lost −41, open-compute −569, and Meldwork −16 (10-24 observed). Across 6,865 consecutive observations of the full repo set that day, 98 repos lost more than 20 stars (1.4%), versus 0.1%–0.4% on a normal day. It looks like GitHub batch-cleaned suspicious accounts that day, and this repo was in the batch being cleaned.
- Curve shape: no spike of traffic lines up with it, yet it rose steadily by +20~60 per day, with a full three-week stall at 0 (like a gap between a package running out and renewal), then steady growth resumed from 10-01. Repos that catch on through content usually show a sharp peak with a long tail, not this shape.
- Our suspect list is empty for it. The reason is that its fork-to-star ratio (10.7%) is close to the median for similarly new repos (about 11%). This boosting service inflates both stars and forks, which sidesteps the "almost no forks" rule.
3. Exposure channels (all follow-on, not sources)
| Channel | Evidence | Nature |
|---|---|---|
| LinkedIn "GitHub AI Agent Trends" leaderboard post | Dr. Wei Liu's post lists "782 stars (+391 this week)" (around early September) | Automatically ranked by weekly star gain; follows star count |
| skillsllm.com / theagenticleaderboard.com / agenticmart.org / githubtree | Listing page, and comparison pages against repos like superpowers | Automated GitHub-scraping directory site |
| Instagram @github_dev Reel | "One App, Four Livestream Agents", with #Mac in the caption (this software only supports Windows) | Automatically generated reposts |
| Arabic-language Facebook page slyverex, YouTube Short "Automate Your Entire Livestream Business With AI" | Search results | Reposting |
| X | x-radar searched the repo name over the past 90 days: 0 posts | None |
| Media / HN / Reddit / Xiaohongshu / Douyin | Searches found no dedicated coverage or discussion | None |
We found no promotional posts from the author, and no discussion of it in Chinese livestream-commerce circles (WeChat Official Accounts, Zhihu, Xiaohongshu).
4. Competitors
| Name | Where | Launched | Scale | How it differs |
|---|---|---|---|---|
| Chanmama (AI review, high-light talk scripts) | Web SaaS | Since 2019, AI review has been added in recent years | Douyin e-commerce data leader, paid membership | Chanmama is itself a data source. It already does AI review and highlight-script extraction on its own; this project just scrapes Chanmama pages and does its own transcription and alignment |
| Oceanengine Baiying / Douyin e-commerce compass / Live Service Platform "post-broadcast replay review" | Big-company features | Douyin official | Every merchant has it | The official backend has replays, minute-level traffic, and highlight moments; this project's review takes Baiying's minute-level traffic table as input |
| Ai Fupan, Changuanjia, Feigua, Douchacha, Huitun | Web SaaS | Mostly 2020–2024 years | Commercial paid tools, aimed at brands and outsourced operators | These are mature paid products for "screen recording + people/goods/venue review"; this project is open source, runs locally, and brings its own key (DashScope/OSS) |
| Feishu Multidimensional Table 'AI Review Master' | Big-company features | 2025 (Jiao Ge Ren Shuang 11 case) | Big-tech ecosystem | Upload a screenshot and get a review out, a lower barrier to entry |
| Tike AI Agent (livestream clip-and-assist Agent) | Windows (Microsoft Store) | 2025–26 | Small | Leans toward real-time in-stream assistance; this project leans toward post-broadcast review |
| Kuaidou Live Recorder (aitoolwang) | Commercial Windows software | — | Listed on Lenovo Store and Sspai | This project's "add to Kuaidou recording" calls this third-party EXE, so the two are upstream and downstream |
| ihmily/DouyinLiveRecorder | Open source | 2023-07 | 11,180 star | Recording only, no AI analysis |
| NanmiCoder/MediaCrawler, Evil0ctal/Douyin_TikTok_Download_API | Open source | 2023 / 2021 | 70K / 21K star | Infrastructure for Douyin data collection and downloads |
| xixihhhh/clipforge, nexscope-ai/eCommerce-Skills | Open source | 2026 | 1.2K / 1.3K star | Also e-commerce AI, but it does product-selling short video generation and cross-border e-commerce skills, not live-stream review |
In the open-source world, the only one found that builds "Douyin live-stream review" into a complete Agent workstation. But commercial SaaS and the official backend have long covered this need, and they hold the data sources.
5. Why it "broke through"
- Stars mainly come from paid star/fork services (evidence in section 2; fairly high confidence, but no stargazer list was obtained).
- Clever topic: "Douyin live e-commerce + AI Agent + local-first + bilingual README", fitting the standard packaging of the Chinese AI Agent library in 2026; aggregator sites and LinkedIn chart posts pick it up automatically, giving secondary exposure.
- Little real use: 207 downloads, 3 real-person issues, and the author has done zero maintenance for over two months.
6. Differentiation
On the product side, the differentiation is that it chains four steps (find creators, transcribe, align script × minute traffic, short-video directing) into one local desktop workstation. In open source, nobody has really done this, and it fits the real workflow of operations teams. But this differentiation is not why it got stars: real users barely showed up, and the star source points to inflated numbers.
7. Takeaways
- The analysis granularity of "script × minute traffic alignment" is a real need. Chanmama's own tutorial says, "the core of the crude method is verbatim transcripts + second markers." A version for TikTok Shop / WeChat Channels / Xiaohongshu live streams, or a SaaS, has room, but it would compete head-on with Chanmama and the official backend.
- Warning: "Chinese vertical Agent + bilingual README + one-off release + bought stars" has become an established playbook at scale (the same account pool serves dozens of libraries). When checking stars on a new Chinese repo, inspect forker account registration dates and fork lists, and check release download counts.
8. Uncertain
- No stargazer list available: the stargazers endpoint returns 404 for this repo, so star accounts cannot be verified directly. The inference above relies on fork accounts, the curve shape, and 10-23 synchronized star drops.
- Unclear how many of the roughly 770 stars in the first two weeks came from real people (e.g., a Chinese community reposting it).
- The author's identity and motive are unclear: outsourced delivery, a job-portfolio piece, or traffic for Kuaidou or some operations company are all possible.