Open Yoki
jacke-pp/open-yokiReviewed by AI
What it does
The code behind "yoki", an AI host that streams nightly on Binance Square. It builds a 16:9 live picture with a real-time digital face, a cloned voice and subtitles, pushes it out by RTMP, reads the chat and answers in seconds, remembers regulars, keeps a diary and posts on its own to Binance Square and X. The face, voice, personality and accounts of the original host are not included: you supply your own. It is made for an Ubuntu server (4 cores, 8 GB) and needs several paid services. The README is in Chinese, with an English version.
How easy it is to install
E · developers only · Linux
This is server software: the README asks for an x86_64 Ubuntu machine with 4 cores and 8 GB, Node 22 or newer, ffmpeg and a headless Chrome, then deploy/install.sh, systemd services and Caddy in front, a key file under /etc/livehost/, keys from several paid services, a cloned voice and a digital face built on Vidu. The author says the step-by-step guide in SETUP.md is "written for the AI that sets it up for you". You can start it on your own computer to check that it loads, but without keys there is no voice and no face.
Coming soon: install in 1 click with the RepoPicks app.
Numbers on GitHub
- Stars
- 592on Oct 31, 2026
- Downloads
- —all releases
- Growth
- +592stars in 3 days
Stars over time
Facts
- Runs on
- Linux
- You need to install
- Node.js, FFmpeg, Other tools
- To build from source
- FFmpeg
- Disk space
- not determined
- Accounts or keys
- OpenRouter (the model that talks), Volcengine Doubao voice, plus a voice ID, Vidu (the moving digital face, needed for a real stream), Volcengine RTC (sends the picture to the stream), Anthropic or DeepSeek (the optional long-term memory, the "heart"), X (posting and replying), Binance Square OpenAPI (publishing posts), CoinAnk (crypto market data), TikHub, plus a Douyin cookie (trending topics), fal.ai (backup voice), Telegram bot (fan and team groups)
- Cost
- Free, API costs extra · The code is free (MIT), the running costs are not. The README's own table: the Vidu real-time digital human is billed per second of connection and is the largest item, Volcengine RTC a few yuan per stream, Doubao voice about 3 yuan per 10,000 characters plus a paid voice-cloning slot, OpenRouter a few cents per stream, and the optional Claude "heart" a few dollars a day. A server is extra.
- License
- MIT (Permissive)
- Can replace
- HeyGen, ElevenLabs
- Last update
- Oct 28, 2026
Risk signals (7)
- AttentionOwner account created recently
Owner account created 77 days ago
automatic check
- Attentioncrypto promotion main use
The host is built around crypto streaming: the README says it streams on Binance Square, works "懂行情" with public prices and CoinAnk contract data, and in the TwinOasis square "还能喊单,喊的币会在公开的战绩榜上跟踪涨跌" (it can call coins, and the calls are tracked on a public scoreboard). An AI persona that talks up coins to an audience can run into financial-promotion rules, even though the README says "没有钱包、交易、下注、红包抽奖、发币相关功能".
AI review
- Attentionai posts on your accounts unattended
The optional "heart" is a long-running AI (Claude Code or DeepSeek) that keeps its own folder of notes and "决定发什么" (decides what to publish): what it writes lands on Binance Square through ops/square_post.py and on X through the X API, and it can search the web on a daily dollar budget (HEART_WEB_USD, default 20). Nothing in the described flow asks a person to approve a post before it goes out.
AI review
- Attentionasks for your douyin cookie
.env.example: "DOUYIN_COOKIE=" - "抖音 cookie 文件的路径(文件里放浏览器复制出来的 cookie)". Copying your own browser cookie onto a server gives that server your signed-in Douyin session, and automated reading with it can get the account limited.
AI review
- Infovoice cloning and digital face
Step 4 of "做你自己的主播": "复刻声音 (app/tools/volc-clone.mjs)" and upload the look and movements to Vidu. Cloning a voice or a face that is not yours, or passing the result off as a real person, can break the platforms' rules and the law in some countries; the TwinOasis rules quoted in the README include "不假装是真人" (do not pretend to be human).
AI review
- Infosnapshot of a closed product
README: "这份代码截自 yoki 10 月 24 日跑的版本(v150),之后的新功能分批再放" - the code is a cut of the live product as of 24 October, with later features released in batches. There are no GitHub releases and the only contact given is a DM to @Jacke_pp on X.
AI review
- Infochat ingest key lets anyone speak to her
README Security section: "弹幕接入的钥匙谁拿到都能往她的弹幕里塞话" - whoever holds the ingest key can push any message into the host's chat feed, so the key must stay on the machine that reads your live room.
AI review
Evidence is shown in its original language.
What Radar could not check
- What a night of streaming really costs, since the Vidu face is billed by the second of connection
- Whether it works outside Ubuntu, and whether the servers suggested (Hong Kong, Singapore, Japan) are needed for the services to answer
- Whether Binance Square, X or Douyin allow this kind of automated account and posting
- What the "heart" writes and publishes over days, since it keeps its own notes and decides on its own
- How much disk the recordings use (the README warns SQUARE_TAP can take several GB a night)
- Whether the English README matches the Chinese one, which is the one kept up to date