Godogen:一句话生成可玩游戏,Claude Code + Codex 的自动游戏开发流水线

Godogen: Describe a Game, Get a Playable Game — Autonomous Dev Pipeline with Claude Code and Codex

Tech-News #Godogen#Claude Code#Godot#Bevy#AI游戏开发#自动化#开源#独立游戏#Tech-News
更新于
🇨🇳 中文

BLUF:Godogen 不是游戏,而是生成游戏的生成器。你描述一个想法,它端到端地交付一个完整的 Godot 或 Bevy 项目——包括代码、素材、场景、自我修复。3200+ Star,2026 年 2 月上线,是目前最完整的 AI 自动游戏开发开源方案。


是什么?

Godogen 是一个基于 Claude Code 或 Codex 的自主游戏开发框架。核心逻辑:

你描述一个游戏想法 → Godogen 自动规划架构、生成素材、编写代码、运行引擎、截图检查、修复画面和逻辑问题 → 最终交付一个完整的 Godot / Bevy 游戏项目。

它不是普通代码生成,而是完整的 AI 驱动游戏开发流水线,像一个”自动游戏工作室”。

Godogen Demo 视频

点击观看 Demo 视频 — 从一句描述到可运行游戏全流程

GitHubhtdt/godogen · ⭐ 3,235 · Python · 2026-02 发布


核心能力拆解

双引擎支持

引擎语言特性
Godot 4C# / .NET 9完整场景树、场景构建器、脚本、资源组织;支持 Android APK 导出
BevyRustCode-first 场景、本地 Bevy 文档查询、确定性截图、最终证明包

选择 C# 而非 GDScript 的原因:docs/gdscript-vs-csharp.md

多模型素材生成

素材类型使用模型
精确参考图、角色图Gemini(Google AI Studio)
纹理、简单物体xAI Grok
图像转 3D 模型Tripo3D
动画精灵(循环检测)Grok Video Generation

截图驱动的自我修复

这是 Godogen 最关键的设计决策:Agent 通过截图判断进度,而不是看代码能否编译

“可见缺陷(穿模、比例错误、动作卡顿、素材缺失)驱动下一轮迭代,而不是被合理化掉。”

这意味着它能修复代码正确但视觉错误的问题——这是普通代码生成工具做不到的。

发布架构

godogen repo(本 repo)
    ↓ publish.sh
game repo(全新仓库,含 CLAUDE.md / AGENTS.md + skills)
    ↓ Claude Code 或 Codex 在 game repo 内运行
实际游戏

Godogen 本身不跑游戏,它生成一个”游戏生成仓库”,Agent 在那个仓库里工作。


快速开始

1. 环境准备

# Godot 项目需要
# - Godot 4(.NET 版本)加入 PATH
# - Python 3 + pip

# API Keys(按需配置)
export GOOGLE_API_KEY=...   # Gemini 图像生成
export XAI_API_KEY=...      # Grok 图像/视频生成
export TRIPO3D_API_KEY=...  # 3D 模型生成

系统依赖:vulkan-toolsxvfbffmpegimagemagick(详见 setup.md

支持平台:Ubuntu、Debian、macOS

2. 发布游戏仓库

# 选择引擎 + Agent 类型
./publish.sh --engine godot --agent claude --out ~/my-game
./publish.sh --engine godot --agent codex  --out ~/my-game
./publish.sh --engine bevy  --agent claude --out ~/my-game
./publish.sh --engine bevy  --agent codex  --out ~/my-game

# 覆盖已有目录
./publish.sh --engine godot --agent claude --out ~/my-game --force

3. 让 Agent 生成游戏

进入 game repo,用 Claude Code 或 Codex 描述你的游戏想法:

# 在 ~/my-game 目录下
claude   # 启动 Claude Code
> 帮我做一个自上而下视角的太空射击游戏,玩家控制飞船,有3种敌人,可以收集能量道具

Agent 接管后自动完成:规划 → 编码 → 资源生成 → 运行 → 截图检查 → 修复 → 完成。

4. 服务器运行(长时间任务)

# 使用 tmux 保持会话
tmux new -s godogen
./publish.sh --engine godot --agent claude --out ~/my-game

# 安装 tg-push,自动将最终视频推送到 Telegram
# https://github.com/htdt/tg-push

对独立开发者和小团队的价值

Godogen 最适合这几类场景:

  1. 快速验证游戏 idea:不用搭环境、不用写 boilerplate,直接描述看效果
  2. 原型加速:有了 Godot/Bevy 基础的开发者,用 Godogen 跑出第一版原型,再手动精修
  3. 学习游戏开发:观察 Agent 如何组织场景树、管理资源、结构化代码,是非常好的学习素材
  4. 小团队 MVP:2-3 人团队可以并行跑多个 game repo 探索不同方向

局限性

  • 完整生成可能需要数小时(推荐 GPU 服务器加速)
  • 复杂游戏逻辑仍需人工精修
  • 依赖多个付费 API(Gemini、Grok、Tripo3D)

与同类工具对比

工具方式引擎自我修复开源
GodogenAgent 驱动全流程Godot + Bevy✅ 截图驱动
Unity AI Muse辅助生成Unity
GitHub Copilot代码补全通用
GameGen 系列视频/图像生成游戏部分

Godogen 是目前唯一将”截图驱动自修复”引入开源游戏自动开发流水线的项目。


源码github.com/htdt/godogen
Demo 视频youtu.be/eUz19GROIpY
作者@alex_erm


© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用,须注明作者姓名及原文链接,不得去除署名后以原创发布。

🇬🇧 English

BLUF: Godogen is not a game — it’s a generator that generates games. Describe an idea, and it delivers a complete Godot or Bevy project end-to-end: code, assets, scenes, and self-repair. 3,200+ stars, launched February 2026, and the most complete open-source autonomous game development solution available.


What Is It?

Godogen is an autonomous game development framework powered by Claude Code or Codex:

You describe a game idea → Godogen plans the architecture, generates assets, writes code, runs the engine, checks screenshots, fixes visual and logic problems → delivers a complete Godot / Bevy game project.

It is not a code generator. It is a full AI-driven game development pipeline — an autonomous game studio.

Godogen Demo Video

Watch the demo — from one sentence to a running game

GitHub: htdt/godogen · ⭐ 3,235 · Python · Released Feb 2026


Core Capabilities

Dual-Engine Support

EngineLanguageFeatures
Godot 4C# / .NET 9Full scene trees, scene builders, scripts, asset organization; Android APK export
BevyRustCode-first scenes, local Bevy docs lookup, deterministic captures, proof bundles

Multi-Model Asset Generation

Asset TypeModel Used
Reference images, charactersGemini (Google AI Studio)
Textures, simple objectsxAI Grok
Image-to-3D modelsTripo3D
Animated sprites (loop detection)Grok Video Generation

Screenshot-Grounded Self-Repair

The most critical design decision in Godogen: the agent judges progress from captured screenshots, not from whether the code compiles.

“Visible defects — clipping, wrong scale, frozen motion, missing assets — drive the next iteration instead of being rationalized away.”

This means it can fix problems that are code-correct but visually wrong — something ordinary code generators cannot do.

Publish Architecture

godogen repo (this repo)
    ↓ publish.sh
game repo (fresh repo with CLAUDE.md / AGENTS.md + skills)
    ↓ Claude Code or Codex runs inside the game repo
Actual game

Quick Start

1. Prerequisites

# API keys
export GOOGLE_API_KEY=...   # Gemini image generation
export XAI_API_KEY=...      # Grok image/video generation
export TRIPO3D_API_KEY=...  # 3D model generation

System deps: vulkan-tools, xvfb, ffmpeg, imagemagick — see setup.md

Tested on: Ubuntu, Debian, macOS

2. Publish a Game Repo

./publish.sh --engine godot --agent claude --out ~/my-game
./publish.sh --engine godot --agent codex  --out ~/my-game
./publish.sh --engine bevy  --agent claude --out ~/my-game
./publish.sh --engine bevy  --agent codex  --out ~/my-game

3. Run the Agent

cd ~/my-game
claude
> Make a top-down space shooter with 3 enemy types and power-up pickups

The agent handles: planning → coding → asset generation → engine run → screenshot check → repair → done.

4. Long Runs on a Server

tmux new -s godogen   # keep session alive over SSH
# Install tg-push to auto-send the final proof video to Telegram

Value for Indie Devs and Small Teams

Godogen is best suited for:

  1. Rapid idea validation: describe and see without environment setup or boilerplate
  2. Prototype acceleration: get a first-pass Godot/Bevy project, then hand-polish
  3. Learning game dev: observe how the agent organizes scene trees, manages resources, structures code
  4. Small-team MVP: run multiple game repos in parallel to explore directions

Limitations: full generation can take hours (GPU server recommended); complex game logic still needs human refinement; depends on multiple paid APIs.


Comparison With Similar Tools

ToolApproachEngineSelf-RepairOpen Source
GodogenFull-pipeline agentGodot + Bevy✅ Screenshot-driven
Unity AI MuseAssisted generationUnity
GitHub CopilotCode completionUniversal
GameGen seriesVideo/image game genPartial

Godogen is currently the only open-source project that integrates screenshot-driven self-repair into an autonomous game development pipeline.


Source: github.com/htdt/godogen
Demo: youtu.be/eUz19GROIpY
Author: @alex_erm


© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.

💬 评论与讨论

使用 GitHub 账号登录后发表评论