Generate a complete, playable 2D game project from a natural language prompt (optionally with reference images). Its AI-native engine writes project state as schema-validated structured text and provides frame-synchronized runtime control, allowing agents to inspect game state, execute actions, and verify behavior themselves. Running on top is an "adversarial team" composed of eight specialized agents, separating implementation from review to discover and fix issues over multiple rounds. Each completed project is also distilled into reusable skeletons, modules, and collaboration conventions for future development (no training required).
Likes 18 · Code Stars 293
Reasons
- The authors claim that on the GameCraft-Bench subset, VibeGame with self-generated art (P2G) consistently outperforms various agent frameworks using the same underlying model, even when those frameworks use a provided asset library.
- Open source and usable: A Phaser-based implementation has been released, including a web dashboard (chat, asset animation preview and boundary adjustment, object properties, playtesting), under Apache-2.0.
- The project page features demos including a civilization-style prototype, a boss fight segment inspired by Hollow Knight, fighting games, roguelikes, and Werewolf.
Usage
- Prerequisites: Python 3.12+, uv, tmux, and either the Claude Code or Codex command-line agent.
- Installation: git clone github.com/tettethu/VibeGame → ./setup.sh (or have your agent read docs/SETUP.md to install).
- Getting started: mkdir my-game && cd my-game → vibegame init → vibegame start, then use /vibegame-build with a goal description in the session; playtest with vibegame run --port auto.
Risks
- It is a framework that orchestrates existing agents and does not include its own model: When using Claude Code, it launches with --dangerously-skip-permissions (skipping permission confirmations), so it is best run in an isolated environment; Codex also requires manually trusting the hooks it writes.
- Currently supports only 2D (Phaser); Godot / Unity / 3D are still on the roadmap; the API costs for the multi-round loop of eight agents are not specified in the README.
- Evaluation is limited to the GameCraft-Bench subset, and conclusions are reported by the authors; the README requests notification to the authors for commercial deployment (not a licensing condition).
Open · Code · Paper Page