A monitor-resident, video-call companion that performs your agent's replies — expression, voice, and a mood that reacts to you — instead of the chat tab you keep open.
Runs on pre-rendered clips — no live inference, the GPU stays free while you work. Fronts Hermes · OpenClaw · anything local.
Your agent replies in text; ghost·vessel splits that reply into three planes and plays it as a living performance — while the model does the thinking.
The agent picks its own expressions per reaction — emotion used as self-expression, not canned playback.
Sinks when you scold it, brightens when you praise it. Idle, it breathes, blinks, and rests its eyes.
Swappable TTS — local (Qwen · MeloTTS · Piper) or cloud (Edge free, ElevenLabs, OpenAI). Cloning supported.
The reply splits into action (emotion) · dialogue (spoken) · data (code & files to the panel, never read aloud).
Everything plays from pre-rendered clips — no live inference holding your card while you work.
An avatar is pure data — clips + persona + theme + voice + emotion map. Drop a folder; it works.
Open-core engine, MIT-licensed. The demo avatar and commissioned work pay for the hours.
Clone the open-core repo, bring your own avatar (or the starter skeleton), point it at your local agent.
Everything, ready to run: the app, Yeoreum's hand-crafted expression set, reactive mood, persona, voice. No install, no Python.
Just the .gvp — her expression set, mood map and voice config, for your own open-core build.
Your character, your persona, your voice — built to spec, delivered as a ready-to-run bundle.
First impressions, feature wishes, bug reports — no login, just say it.