Geny
Multi-agent VTuber + autonomous worker platform that pairs a chatty Live2D/Spine avatar with tool-using sub-workers, orchestrated in a 3D city playground via five LLM backends.
Ficha actualizada 23 sept 2026, 1:25 · upstream 23 sept 2026, 1:18 · release v0.28.9
Qué es Geny
Descripción editorial construida a partir del repositorio, su documentación y sus señales de mantenimiento.
Geny is an open-source multi-agent platform that combines conversational virtual avatars (Live2D/Spine) with autonomous task execution. A VTuber session handles natural language interaction and personality management, while paired sub-workers perform real-world tasks such as file operations, web browsing, or tool invocation. All interactions are mediated through a structured environment manifest that defines every pipeline stage, provider configuration, and tool set in advance. This enables deterministic reproduction of agent behaviors across sessions without code changes.
The platform runs on FastAPI with Next.js 16 frontend powered by React Three Fiber for the 3D city playground. Five LLM backends are supported: Anthropic, OpenAI, Google, vLLM (self-hosted), and Claude Code CLI. Each backend is configured per environment via a dropdown selector in the settings panel.
The architecture separates concerns between the VTuber chat layer, the agent orchestration engine (geny-executor with 21 stages), the sandboxed execution environment, and the GAPT workspace platform for broader devops use cases. Sessions are persisted using PostgreSQL and Redis, with memory stores supporting vector retrieval and structured tool traces.
- Automated software testing: Run unit tests against a live UI while a VTuber narrates test results in Korean/English
- Customer support automation: A VTuber handles customer inquiries while sub-workers process tickets, update knowledge bases, or trigger workflows
- Content creation pipeline: Generate video scripts, edit assets, and render scenes with a VTuber as the creative lead and AI agents handling production tasks
- Data analysis workflow: Run data cleaning, visualization, and reporting pipelines with an agent that explains findings to a human audience in real time
- IT support automation: Diagnose server issues, deploy patches, or update configurations while a VTuber documents the process for onboarding new team members
- Five LLM backends (Anthropic, OpenAI, Google, vLLM, Claude Code CLI) with per-session configuration and health monitoring
- Manifest-driven environment system that allows deterministic reproduction of agent behaviors without code changes
- VTuber-sub-worker pairing model where the avatar handles conversation and the sub-worker executes real tasks through MCP-bridged tools
- 3D city playground visualization using Three.js and React Three Fiber for intuitive agent behavior observation
- Stable error codes (exec.<component>.<reason>) that surface with localized prompts in Korean or English depending on session language
- Self-hosted deployment requires manual containerization setup; Docker Compose is supported but not included in the repo
- No GPU acceleration available — inference runs entirely on CPU, which may become a bottleneck for large model contexts
- VTuber avatar generation and animation are limited to 2D Live2D or Spine formats with no native 3D character creation
- Tool execution is sandboxed via CLI runners; complex toolchains requiring system-level access (e.g., sudo) cannot be used without manual configuration
- Memory usage grows with session history and vector store size — PostgreSQL requires careful tuning for long-running sessions
Perfil de infraestructura
Estimación inicial derivada del repositorio y de los servicios detectados.
Despliegue detectado
Qué señales reales de instalación hemos encontrado.
Install Geny using Docker Compose: `docker-compose up -d`. The app runs on port 3000 (frontend) and 8000 (backend). For production, use a reverse proxy like Nginx with SSL termination. Set environment variables via `.env` file or Kubernetes secrets. Enable auto-update from v0.3.0+ by configuring GitHub Releases webhook in the desktop connector. For self-hosted LLM backends, run them inside GAPT containers within Geny's sandboxed workspace to ensure tool isolation and security.
Stack y dependencias
Componentes inferidos a partir de manifests, configuración y documentación.
Puede sustituir a
Relaciones detectadas por el análisis del repositorio y su categoría.
Experiencia real de la comunidad
Ayuda a contrastar la estimación automática con instalaciones reales.
Todavía no hay datos comunitarios para este proyecto.