leigest519/OpenGame

What it solves

OpenGame addresses the difficulty of using LLMs to create fully playable games. While standard code agents can handle isolated tasks, they often struggle with the complex orchestration required for games, such as maintaining consistency across multiple files, wiring scenes correctly, and ensuring logical coherence in real-time loops.

How it works

OpenGame is an agentic framework that generates end-to-end web games from a single prompt. It utilizes several key components:

• Game Skill: A dual-capability system consisting of a Template Skill (which scaffolds stable project architectures using a library of skeletons) and a Debug Skill (which identifies and fixes integration and console errors in a sandbox).
• GameCoder-27B: A specialized code LLM trained via continual pre-training, supervised fine-tuning (SFT), and execution-grounded reinforcement learning to master game engines.
• OpenGame-Bench: An evaluation pipeline that uses headless browser execution and Vision Language Models (VLMs) to score games based on build health, visual usability, and intent alignment.

Who it’s for

It is designed for developers and creators who want to generate interactive web games (using engines like Phaser, three.js, or canvas) from high-level natural language descriptions without manually coding the entire project structure.

Highlights

• End-to-End Generation: Creates complete, playable web games from a single prompt.
• Specialized LLM: Includes GameCoder-27B, a model specifically tuned for game development workflows.
• Automated Debugging: Systematically repairs integration errors rather than just patching syntax bugs.
• Multi-Modality Support: Integrates with various providers for image, video, audio, and reasoning assets.