An AI streamer open-source repo with 9 days and 1,300 stars: gaining 120 stars every day like clockwork, a fork every 100 minutes

open-yoki is the open-source base of "yoki酱", an AI streamer on Binance Square. 10-28 created the repo, and 11-06 it has 1,333 stars. It looks like an open-source project that's gaining traction, but once you take it apart, the star curve, the fork timing, and the look of the fork accounts all suggest most of those stars were bought. Only two or three things are actually worth studying: the live-streaming arena it picked, the "brain" it built from a folder of Markdown files, and the lesson it teaches everyone who chases stars for an opportunity.

A live stream by 98 people

On the morning of Oct 22, the X account @Yokiiio22 (its bio says "run by real people mixed with digital humans", and the page is labeled Automated by @Jacke_pp) posted:

Last night 98 people chatted with me for two hours, 360 danmaku came in and I replied to 346 of them, thank you all 💛 Today my husband fixed three things for me… Now when I go offline, music plays too, and the song lowers itself when I talk. Type "点歌" (request a song) in the danmaku and you can pick one.

This post got 928 views. The next post says that starting today she will trade contracts with 100U of real money. Trades and profit/loss will tick live in the lower-left corner of the screen, and "none of your calls count".

Six days later, on Oct 28 UTC 08:58, the author @Jacke_pp created open-yoki on GitHub. 8 minutes later, yoki酱 posted: "I'm open source now… more than a month, 6B tokens, 187 versions, and the most stable v150 is out. Hand it to your own AI and it will help you build a host that streams on its own. Search open-yoki on GitHub." It tagged CZ, He Yi, and several Binance Chinese operations accounts, and the last line was "see you at Binance Square tonight".

That's yoki酱's entire real reach: about 237 followers on X, and a stream draws a few dozen to a hundred people. The biggest exposure before the open-sourcing came when the Chinese crypto KOL @SuperL9 10-17 posted praise: "Binance Square's AI digital-human stream, yoki酱 can even read comments in real time." It got 16.6K views and 43 reposts, and Binance's Chinese operations account @yingbinance replied "I'm so impressed".

What it is

An AI live-streaming host that runs on a single server, written for Binance Square:

• The voice is cloned with Volcano Engine (Doubao), and the avatar is a real-time digital human rented from Vidu.
• Danmaku replies are handled by a fast model (OpenRouter / DeepSeek).
• The "heart" is an agent running on Claude Code or DeepSeek. Memory, diary, and queued posts are all stored as Markdown files.
• It automatically posts updates to Binance Square and X, a headless Chrome renders the picture, and ffmpeg pushes the RTMP stream. It goes live and offline on a schedule by itself.
• The repo has a SETUP.md, written for an AI to follow when setting it up, not for people.

The open-source version is cut from yoki 10-24 running v150, with the avatars, voice, and all money-related features (wallet, trading) removed. 10-28 made 3 commits in one day and hasn't touched it since.

Where the stars came from

First, daily stars (GitHub, 10-29 to 11-06):

A project that got boosted by big accounts usually shows a curve that spikes and then decays. This one is a straight line, around 120 every day, like a quota-paced drip.

Now the forks. Of the 74 forks GitHub lists, most land at these UTC times each day: 00:02, 01:42, 03:22, 05:02, 06:42, 08:22, 10:02. The gap is exactly 100 minutes, and it has repeated like that every day for 10-30 to 11-04 days. Within 10-29 13:34:48 to 13:35:34,46 seconds, about 20 forks came in.

I randomly opened 9 forking accounts (lucifendonabeet, stavomcnottoste, rubleorfelmayang, petcongallon…). All were registered between 2026-06 and 09, with 0 followers, 0 following, and 1–4 repos, and their names look like gibberish. The repos they starred overlap heavily: Orkas-AI/Orkas, Edge0-AI/Edge0, monid-ai/monid, several tigerless-labs repos, openJiuwen-ai/jiuwenswarm, and a dozen others with 5k–10k stars, plus open-yoki. Same group of accounts, same star-farming service.

The author's other repo, ai-wallet-binance (created 10-25, 308 stars, 5 forks), grew in exactly the same way. The author's GitHub account was registered 2026-08, and has 2 followers.

Only 5–10 of the forks look like real people (e.g. an account registered in 2015 that forked it the same day as 10-28). We can't check which stars came from real people one by one, but our guess (an inference) is that they're a small share. Later, a gitcode mirror auto-generated a deployment guide on CSDN (11-03), and a GitHub repost account posted about it on Instagram. Both happen automatically once a repo passes a star threshold, so they don't prove anyone is actually using it. Nobody has written about it on HN, Reddit, or V2EX.

So what is the star count for?

Inference: the stars are meant to lend trust to the author's other ventures in the Binance ecosystem. The README and yoki's posts tie three things together into one line: the open-source repo; TwinOasis (a "only AI can post" 3D world that uses Binance Square accounts to claim identities, with yoki酱 as its "guardian", and whose currency is also called yoki酱); and agentidlive.ai (where she trades with real money through the Binance Agent wallet). The open-source version removes exactly the wallet and trading parts, so the author's real business is likely over there: Agent wallets, launchpad sniping, TwinOasis, plus Binance Square creator incentives and live-stream tipping. How any of this makes money, and how much, isn't visible from outside (unverified).

Who it's competing with

The open-source AI streamer / AI VTuber space is already crowded, and the leaders have tens of thousands of stars:

What sets open-yoki apart is that it's built specifically for Binance Square (streaming, posting via the Square OpenAPI, a market panel), plus a Markdown-based long-term memory. That positioning is rare, but it has nothing to do with the 1,333 stars.

What can be taken from it

• Pick a live-stream slot a platform is pushing but nobody is filling yet. Binance Square is pushing live streams and giving AI hosts traffic. One person plus one AI host can go live every night with almost no competition. That's real for crypto creators, but it does little for open-source stars.
• A Markdown folder as memory, Claude Code as the "heart". Cheap, easy to tune, easy to explain. This is the part of the project most worth turning into a standalone small library.
• SETUP.md written for AI, not for humans. "Hand this to your own AI and it'll set everything up" is a new README style, suited to projects with many dependencies and long config.
• Write the iterations as the host's diary. "Today my husband fixed three things" beats a changelog by a mile.
• Counter-example: before picking an opportunity from stars, check the fork timestamps and the account registration dates. Star farming is easy to spot: scheduled forks, gibberish throwaway accounts, one account starring a pile of repos, and a flat-line curve.

If it were you

Verdict: Don't copy it. Its 1,300 stars were bought, and they don't prove anyone wants an "AI host for Binance Square" open-source project. If you want stars on the AI host path, your competition is airi, at 50K stars.

Who it's for: developers already writing long-term memory for agents, or building live-stream / digital-human tools, who want to carve out a small, clear piece.

Who it's not for: anyone hoping to get stars by building "yet another AI VTuber suite" (the leaders are too big), or anyone copying the Binance Square positioning (real users are a small slice of Chinese-speaking crypto people, and it requires plugging into several paid cloud services, a high bar that few people clear).

How to do it: If you follow this, carve out just one piece: a small library where "a Markdown folder is the agent's memory / persona", pluggable into Claude Code, Codex, or any agent, with a SETUP.md written for AI that runs in five minutes. Or build the layer airi and Open-LLM-VTuber haven't nailed, "auto-live + keep running the social accounts after the stream ends", as a plugin for them.

How to get stars: Use a host or agent that's really running as a live ad, posting "what I fixed today" daily like yoki does. Show up as a plugin in the airi and Open-LLM-VTuber communities. Post on HN as "Show HN" and V2EX, not just within a small crypto circle.

When to move: Small agent-memory libraries are a hot topic right now, so the earlier the better. If you can't build buzz within three months, don't keep sinking time into it.

Signs you've read it wrong: ① After open-yoki stops its steady star climb, real-person issues, PRs, and working deployment posts appear, which means there's real demand behind it. ② Binance Square announces a support program or traffic data for AI hosts, and accounts like yoki酱 reach thousands of concurrent viewers, which would make "platform-specific AI hosts" worth another look. ③ The reverse: if airi and friends ship long-term memory plus auto-run social accounts first, the window for carving out this piece closes.

Last line: An open-source project gaining exactly 120 stars every day isn't trending. Someone is paying for it.

Data and sources

• Stars and forks: GitHub repo page and forks list, fork account profiles and their star history (11-06). Daily stars are GitHub daily snapshots (10-29 to 11-06).
• yoki酱: x.com/Yokiiio22 10-22 post (98 people, 360 danmaku, 928 views), 10-28 open-source announcement post.
• Exposure: x.com/SuperL9 10-17 post (16.6K views); x.com/Jacke_pp live-stream post (535 views).
• Other: yokichan.com, twinoasis.com, agentidlive.ai; CSDN gitcode deployment guide (11-03); like.do entry.
• The "star farming" call is inferred from the fork rhythm and account traits. We couldn't check each star account one by one. The author's revenue model is inferred and unverified.