Abstract

Game development is a complex discipline requiring the synergy of human creativity, professional engineering skills, and the integration of game engines and complex state management. While Large Language Models (LLMs) and code agents excel at solving isolated programming tasks, they often struggle to generate fully playable games directly from high-level designs, frequently failing due to cross-file inconsistencies and logical incoherence. To bridge this gap, we present OpenGame, the first open-source agentic framework explicitly designed for end-to-end web game creation. At the core of OpenGame is Game Skill, a novel, reusable capability comprising Template Skill and Debug Skill, which allows the agent to organically learn to scaffold stable project architectures and systematically resolve integration errors. To support this framework, we introduce a specialized foundation model: GameCoder-27B, a Code LLM specifically optimized for game engine mastery and logic via Supervised Fine-Tuning (SFT) and Reinforcement Learning. However, evaluating such agents poses a formidable challenge, as verifying the interactive playability of a game is far more complex than standard static code evaluation. To tackle this, we present OpenGame-Bench, a comprehensive benchmark designed to dynamically assess an agent's capability to construct interactive web games. We hope that OpenGame not only bridges human creativity to playable content, but also pushes code agents from solving discrete software engineering problems to building complex, interactive real-world applications. All frameworks, models, data, and evaluation pipelines will be fully open-sourced.