OpenGame - CUHK's Open-Source End-to-End Web Game Agent Framework

What is OpenGame

OpenGame is the first end-to-end web game agent framework open-sourced by CUHK MMLab, which automatically generates playable web games through natural language prompts. The framework's core Game Skill consists of Template Skill (a project skeleton template library) and Debug Skill (a dynamic debugging protocol), equipped with GameCoder-27B, a three-stage code model trained specifically for game engines, and paired with the OpenGame-Bench dynamic evaluation benchmark, systematically solving the challenges of cross-file inconsistency and logic defects.

OpenGame's Main Features

• End-to-End Game Generation: Accepts natural language design requirements and automatically outputs a complete, buildable, and runnable web game project, covering the entire process from concept to finished product.
• Game Skill System (Game Skill)
  • Template Skill: Based on physics mechanism classification (such as side-scrolling gravity, top-down continuous movement, discrete grid, etc.), automatically selects and instantiates a stable project template family, ensuring cross-file structural consistency.
  • Debug Skill: Maintains a dynamic debugging protocol, accumulating error signatures, root causes, and verified fix solutions, achieving systematic integration error fixes rather than isolated syntax patches.
• Game Design Document (GDD) Auto-Generation: Transforms natural language requirements into a technical GDD, dynamically loading engine API constraints to ensure mechanisms are feasible under the selected framework.
• Multimodal Asset Generation: Calls image, audio, and video generation models to automatically synthesize backgrounds, character animations, sound effects, tilemaps, and other resources based on the GDD asset list.
• Three-Tier Reading Strategy Code Implementation: Through progressively loading API summaries, target source code, and implementation guides, combined with the template method pattern (Hook-Driven) to inject game logic, avoiding context overflow.
• Verification and Self-Correction Loop: Executes builds and tests through a headless browser, parses compilation output, and iteratively fixes until the game is playable.

OpenGame's Technical Principles

• GameCoder-27B Three-Stage Training: Based on the Qwen3.5-27B backbone, through continuous pre-training (CPT, Phaser/JS game code and documentation), supervised fine-tuning (SFT, GPT-Codex generating complex design prompts and MiniMax synthesizing high-quality solutions), and execution reinforcement learning (RL, single-file module-level unit test feedback), establishes game engine architecture priors and API usage patterns.
• Physics-First Classification Mechanism: Classifies game requirements into prototypes based on physical constraints and spatial mechanisms (such as "falling without ground support" mapped to the platformer prototype), guiding subsequent template selection.
• Template Method Pattern (Hook-Driven Implementation): The agent copies template files and overrides specified hook methods (such as setupCustomCollisions), injecting specific logic while preserving the determinism of basic lifecycle management.
• Dynamic Agent Evolution: Template Skill evolves from a single meta-template M0 through accumulated experience into a template library L containing five families (gravity side-view, top-down continuous movement, discrete grid, path wave, UI-driven). Debug Skill records error signatures, root causes, and verified fixes from execution failures, forming a Living Debugging Protocol P, which includes pre-execution validation for high-frequency inconsistency categories.
• OpenGame-Bench Dynamic Evaluation: Generates games by executing them through a headless browser, combined with VLM (Vision-Language Model) judging, scoring across three dimensions—Build Health, Visual Usability, and Intent Alignment—surpassing static unit tests.

How to use OpenGame

• Environment Setup: Install Node.js 20+, clone the repository and run npm install、npm run build、npm link, register the opengame command to the system path.
• Configure API Keys: Set OPENAI_API_KEY and optionally OPENAI_BASE_URL、OPENAI_MODEL; also configure multimodal provider keys (image, video, audio, reasoning), such as DashScope, Doubao, etc., refer to .env.example。
• Generate a Game: Create an empty directory, run opengame -p "Build a Snake clone with WASD controls and a dark theme." --yolo, the agent will automatically complete the end-to-end build.
• Run Locally: Enter the generated project directory, run npm install and npm run dev, open http://localhost:5173 in the browser to play.
• Advanced Configuration: Behavior can be adjusted via settings.json(user-level ~/.qwen/settings.json or project-level .qwen/settings.json) and CLI parameters; setting GAME_TEMPLATES_DIR and GAME_DOCS_DIR allows creating games anywhere using absolute paths.

Key information and usage requirements of OpenGame

• Open Source License: Open source project, built as an extension based on qwen-code.
• Operation Mode: Currently in command-line Headless mode, one-click prompt-driven generation.
• Model Dependency: Supports OpenAI-compatible APIs, with optional backends such as GameCoder-27B (locally deployed) or Claude Sonnet 4.6.
• Multimodal Requirements: You must configure API keys for image, video, and audio generation services yourself; the framework does not include default providers.
• Node.js Version: Requires Node.js 20 or above.