Godogen: Autonomous game development for Godot and Bevy with Claude Code and Codex

Watch the demos · Prompts

Describe a game. Godogen plans it, writes the code, generates assets, runs the engine, checks screenshots, and fixes what looks wrong.

This repo is not a game. It is the source for a generator that produces games: godogen -> game repo -> game. You publish the skills into a fresh game repo, choosing the engine and host-agent flavor, then the agent runs inside that repo to build the actual game.

Source layout

The source is organized along the engine axis:

• shared/ — engine-agnostic godogen stages, asset-generation tooling, shared stop hook, and common game-repo instructions
• godot/ — Godot-specific godogen stages, Godot capture helpers, and the godot-api skill
• bevy/ — Bevy-specific godogen stages, Bevy capture helpers, and the bevy-help skill

Claude Code vs Codex is a publish-time render choice, not a separate source tree. The root publish.sh renders the right runtime layout for the chosen engine and host agent.

What skills do

• Godot 4 output — real C#/.NET projects with proper scene trees, scene builders, scripts, and asset organization.
• Godot Android export — debug APK export remains available when the user requests an Android app.
• Bevy output — Rust/Bevy projects with code-first scenes, local Bevy docs lookup, deterministic capture guidance, and final proof bundles.
• Asset generation — Gemini creates precise references and characters; xAI Grok handles textures and simple objects; Tripo3D converts images to 3D models. Animated sprites use Grok video generation with loop detection.
• C# / .NET 9 for Godot — Godot output uses C#. See why C# over GDScript.
• Frame-grounded self-repair — the agent is carefully prompted to judge progress from captured screenshots, not from code that compiles, so visible defects (clipping, wrong scale, frozen motion, missing assets) drive the next iteration instead of being rationalized away.
• Telegram proof push — published repos install a stop hook that pushes the latest screenshots/result/{N}/video.mp4 to Telegram when tg-push and the TG_* env vars are configured. No-op otherwise.
• Runs on commodity hardware — any machine with the relevant engine toolchain, Python, and the required API keys can run the pipeline.

Getting started

Prerequisites

• Godot 4 (.NET build) on PATH for Godot projects
• Rust/Cargo plus local Bevy docs for Bevy projects
• Python 3 with pip
• API keys as environment variables:
  • GOOGLE_API_KEY — Google AI Studio for Gemini image generation
  • XAI_API_KEY — xAI Grok for image/video generation
  • TRIPO3D_API_KEY — Tripo3D for 3D generation
• System packages from setup.md: vulkan-tools, xvfb, ffmpeg, imagemagick, plus platform-specific extras
• Tested on Ubuntu, Debian, and macOS
• Claude Code or Codex

Publish a game repo

Pick the engine and host agent:

./publish.sh --engine godot --agent claude --out ~/my-game  # CLAUDE.md + .claude/skills/
./publish.sh --engine godot --agent codex  --out ~/my-game  # AGENTS.md + .agents/skills/
./publish.sh --engine bevy  --agent claude --out ~/my-game
./publish.sh --engine bevy  --agent codex  --out ~/my-game

Pass --force to wipe existing contents at the target before publishing — use this when re-publishing over a previous run.

Bevy docs setup

If you're working on Bevy generation, configure and populate a shared Bevy docs folder once after clone:

./setup_bevy_docs.sh /absolute/or/user/path/to/bevy-docs

The setup script links bevy/skills/bevy-help/docs/ to that folder, clones the Bevy docs sources, and builds local rustdoc for the current stable release. No default path is assumed. See setup.md for the full workstation setup.

Running on a server

A full generation run can take hours, so it's convenient to offload it to a server, ideally a GPU instance, since engine rendering and video capture are much faster with hardware acceleration.

• Keep the session alive across SSH drops with tmux or screen.
• Install tg-push: the stop hook auto-sends the final proof video to Telegram on completion.
• Enable remote control so you can check in and steer the run from any device — both Claude Code and Codex have official remote-control interfaces.

Improving the skills

After a full generation session, ask the agent you used to review how the pipeline performed:

Analyze this session. Were the instructions optimal? Flag anything that was too obvious, missing, or misleading. Did any tools pollute context with noise? Did the capture loop catch the real problems? Any tool failures or workarounds?

Changelog

See CHANGELOG.md.

Follow progress: @alex_erm