jacke-pp/open-yoki

open-yoki is a self-hosted AI livestream host: a headless Chrome draws a 16:9 frame, a real-time digital avatar speaks in a cloned voice, and ffmpeg pushes the result over RTMP to Binance Square or any RTMP endpoint. It is aimed at people who want a customised virtual presenter rather than a finished consumer product.

What you get and how it works

The host reacts to live comments in a few seconds using a fast chat model (DeepSeek via OpenRouter by default), with mood, gestures and tone. Optional "heart" modules — a long-running Claude Code CLI or DeepSeek process living in a folder of Markdown files — keep memories, a diary and impressions of regulars, and decide what to post. A big-move mechanism plays pre-made short clips when the dialogue calls for them, and an optional radio mode runs between streams with voice over a still image.

The server itself does not log in to any platform. Something that reads your live room — your own script or a platform bot — posts events to local ingest endpoints for comments and room events such as entering, following, gifts, tips and stream end. A generated key in the data directory authenticates those calls, and an outbox API returns scheduled posts for a separate poster script to publish. A "town" integration lets an external AI world read the host's public card and mood and send events back.

Before you start

The repository explicitly excludes the original host's face, voice, persona, songs, sound effects and clips; the bundled "Lili" example is a placeholders only, including a plain colour-gradient avatar. You supply your own persona, portrait, cloned voice and media assets. There are no accounts, keys or domains bundled, and no wallet, trading, betting or red-packet features; market talk uses public prices only.

Running the full stream requires paid third-party services: Vidu for the animated face (the largest cost, billed by connected seconds), Volcengine RTC for delivery, Doubao voice cloning for speech, and OpenRouter or DeepSeek for dialogue. Optional X, CoinAnk, TikHub and Telegram integrations add their own accounts or fees. The local trial needs Node.js 22+ and ffmpeg; without API keys the brain and voice stay off and the avatar is a static image. The suggested server is x86_64 Ubuntu, 4 cores and 8 GB, in Hong Kong, Singapore or Japan. A detailed setup guide is provided for an AI assistant to follow.

Operational cautions matter: whoever holds the comment-ingest key can inject lines into her stream, so keep it only on the machine reading your live room; the web live room listens on 127.0.0.1 only and should be put behind Caddy when exposed. Keys belong in root-owned environment files, not in the host config.

Licence

The code is MIT; bundled third-party files keep their own licences.

Project source and documentation