AI-powered game creation platform. Describe your game in plain English, watch an AI agent team build it collaboratively with live preview. Supports 2D games with Phaser 3 and 3D games with Three.js.
GameForge is a learning repo — it demonstrates how to build a multi-agent AI pipeline using the Claude Agent SDK and Google Gemini (Nano Banana for sprites, Lyria for music) to orchestrate a team of specialized AI agents that collaborate to produce working Phaser 3 (2D) and Three.js (3D) games.
- Claude Agent SDK patterns — programmatic agent control, subagents, structured tool use
- MCP tool servers — building custom Model Context Protocol tools for agent capabilities
- Multi-agent orchestration — designing pipelines where agents hand off work to each other
- Google Gemini Nano Banana — generating game sprites and backgrounds from text descriptions
- Google Lyria — composing background music via the RealTime WebSocket API
- Phaser 3 + Three.js game development — scaffolding, code generation, and live preview for 2D and 3D games
- Full-stack TypeScript — Nx monorepo, Next.js, Express, WebSocket, Zustand
When a user describes a game in the chat, the orchestrator runs a sequential pipeline of specialized agents:
User Message
│
▼
┌─────────────┐ Creates a Game Design Document (GDD)
│ Designer │ with mechanics, art direction, and audio specs
└──────┬──────┘
│
▼
┌─────────────┐ Generates sprites and backgrounds via
│ Artist │ Google Gemini Nano Banana (runs if GDD has artDirection)
└──────┬──────┘
│
▼
┌─────────────┐ Composes background music via Google Lyria
│ Musician │ RealTime API (runs if GDD has musicDirection)
└──────┬──────┘
│
▼
┌─────────────┐ Writes Phaser 3 or Three.js TypeScript code
│ Developer │ using the GDD and generated assets via MCP file tools
└──────┬──────┘
│
▼
┌─────────────┐ Launches the game in Playwright, captures
│ QA │ screenshots, reports bugs back to the user
└──────┬──────┘
│
▼
User Feedback → loops back to Designer for iteration
Each agent uses the Claude Agent SDK with custom MCP tool servers. The orchestrator manages conversation context so agents can see prior work and iterate based on QA feedback.
| File | What It Demonstrates |
|---|---|
apps/orchestrator/src/agents/teamOrchestrator.ts | Pipeline orchestration — how agents are sequenced and handed context |
apps/orchestrator/src/tools/gameToolServer.ts | MCP tool server — custom file tools that agents use to read/write game code |
apps/orchestrator/src/agents/prompts/ | Agent system prompts — how each agent role is defined |
apps/orchestrator/src/assets/assetGenerator.ts | Gemini Nano Banana integration — text-to-image generation for game sprites |
apps/orchestrator/src/music/musicGenerator.ts | Lyria RealTime API — WebSocket streaming for music composition |
apps/orchestrator/src/tools/assetToolServer.ts | MCP tools for asset generation — generate_asset, get_asset_status |
apps/orchestrator/src/tools/musicToolServer.ts | MCP tools for music generation — generate_music, get_music_status |
docs/research/ | API research notes — Claude Agent SDK, Gemini Nano Banana, Lyria, Phaser, Three.js |
apps/
studio/ Next.js 16 frontend (port 4001) — chat + game preview
orchestrator/ Express + WebSocket backend (port 4000) — AI agent pipeline
packages/
shared-types/ TypeScript types shared across apps (messages, sessions, agents)
game-templates/ Phaser 3 and Three.js starter templates for scaffolding
sessions/ Runtime game project data per session (gitignored)
- Node.js 20+
- npm 10+
- An
ANTHROPIC_API_KEYfor Claude (required) - A
GOOGLE_AI_API_KEYfor asset/music generation (optional)
# Install dependencies
npm install
# Copy environment variables
cp .env.example .env.local
# Edit .env.local and set ANTHROPIC_API_KEY (and optionally GOOGLE_AI_API_KEY)# Build shared packages
npx nx run-many -t build --projects=shared-types,game-templatesRun the orchestrator and studio in separate terminals:
# Terminal 1 — Orchestrator backend
npx nx serve orchestrator
# Terminal 2 — Studio frontend
npx nx dev studioThen open http://localhost:4001 in your browser.
# Start all services
docker-compose up
# Start with E2E test runner
docker-compose --profile e2e upHealth checks are configured for both services. The studio waits for the orchestrator to be healthy before starting.
| Variable | Required | Description |
|---|---|---|
ANTHROPIC_API_KEY | Yes | Claude API key for the AI agent pipeline |
GOOGLE_AI_API_KEY | No | Google AI API key — enables Artist (Nano Banana) and Musician (Lyria) agents |
ORCHESTRATOR_PORT | No | Orchestrator HTTP/WS port (default: 4000) |
STUDIO_PORT | No | Studio frontend port (default: 4001) |
NEXT_PUBLIC_ORCHESTRATOR_URL | No | Studio → orchestrator REST URL (default: http://localhost:4000) |
NEXT_PUBLIC_ORCHESTRATOR_WS_URL | No | Studio → orchestrator WebSocket URL (default: ws://localhost:4000) |
Tests are mandatory for all code. Run them with:
# All tests
npx nx run-many -t test# Single project
npx nx test studio
npx nx test orchestrator
npx nx test shared-types
npx nx test game-templates
# Type checking
npx nx run-many -t typecheck# Build all projects
npx nx run-many -t build
# Build specific project
npx nx build orchestrator
# View the project dependency graph
npx nx graph- Frontend: Next.js 16 (App Router), Tailwind CSS, Zustand
- Backend: Node.js, Express, WebSocket (ws), esbuild
- Game Engines: Phaser 3 (2D), Three.js (3D)
- AI: Claude Agent SDK (TypeScript), Claude Opus
- Asset Generation: Google Gemini Nano Banana (sprites, backgrounds)
- Music Generation: Google Lyria RealTime API (background music)
- Build: Nx monorepo, Vite (game projects)
- Testing: Vitest (unit/integration), Playwright (QA + E2E)
docs/research/— API research notes on Claude Agent SDK, Gemini Nano Banana, Lyria, Phaser 3, Three.jsdocs/demo-walkthrough.md— Step-by-step walkthrough of a demo session
This project was built iteratively in phases. See tasks/todo.md for the full phase-by-phase development log, and tasks/lessons.md for lessons learned during the build.