VGAI — build and adapt games with AI agents

VGAI is an AI-native game engine, visual editor, and universal host. A human or coding agent can create a new game, continue a VGAI project, or bring an existing Three.js, React Three Fiber, PixiJS, or React game without treating the engine repository as the game project.

What do you want to do?

Create a new game

Open the web editor for a zero-install start, or use the one-command local setup. Both routes open the visual editor; local projects also include agent skills and MCP configuration.

Start from an example

Open a working example immediately, then inspect it in Play mode or scaffold an editable standalone copy with the CLI.

Continue a VGAI project

Run npm run dev in the project folder. It opens or reuses that project’s editor session; npm run game is the separate player-only surface.

Bring an existing game

Begin with the adapter and ingest guide. VGAI detects the renderer and capability tier before adding metadata, and never silently rewrites the original game.

Use the engine as a library

Install @vgai/engine directly when you want its TypeScript APIs inside your own project rather than the full editor-first workflow.

Learn or contribute

Choose Get Started and Foundations for the learning path, the Manual for concepts, or the Reference for exact APIs and schemas.

One project, every surface

Browser, terminal, coding agent, and local editor are entry surfaces—not different game formats. Authored games use the same vgai.game.json, explicit adapter roots, ordinary source files, and Play-mode verification.

Explore by area

Editor Guide Viewport, gizmos, inspector, asset browser, prefabs, build panel.

React UI Authoring Author declared React roots through the hierarchy and JSX source round-trip.

Loading Existing Games Bring unmodified Three.js games into the editor via adapters.

Multiplayer Server-authoritative games with Colyseus rooms + schema sync.

Tooling & Publishing CLI, scaffolding, hot reload, web builds.

Quick Starts Short, focused build-one-thing guides.

The learning structure follows the proven separation used by established game engines—tutorials, focused how-to guides, conceptual manual, examples, and API reference—while the project and agent flow is designed for AI-first authoring.