Game-Creation-Agents (GCA) is an end-to-end pipeline that turns a natural-language idea into a playable, verified browser game. The creation flow is:
- [2026-10-14] We are pleased to announce the initial public release of Game-Creation-Agents (GCA), an end-to-end agentic pipeline for designing, building, and verifying playable browser games.
Requirements:
- macOS or Linux; Windows users should use WSL2;
- Python 3.11 or newer;
uv;- Node.js 20 or newer and npm;
- Playwright with Chromium;
curl,ffmpeg, andffprobe;- Claude Code with project MCP support;
- an OpenAI API key for image generation and editing;
- a Volcengine ARK API key for Seedance video generation and Doubao VLM;
- no background-removal model: the default
assets.matting.backend: color_keymattes by pure computation against the generated staging background. An optional RMBG-2.0 matting service can be opted into for soft-alpha edges โ see Background removal.
The current preflight checks provider credentials, the configured RMBG/color-key
matting routes, browser, media, and filesystem dependencies required by the full pipeline.
Seedance sends local input frames inline, so object storage is not required.
Host binary requirements and the minimum Node.js version are declared in
scripts/preflight-requirements.toml.
Install system prerequisites
On macOS, install Homebrew first, then:
brew install git uv node ffmpeg python@3.11
npm install -g @anthropic-ai/claude-code playwright
npx playwright install chromiumOn Ubuntu / Debian:
sudo apt-get update
sudo apt-get install -y git curl ffmpegThen install uv,
Node.js 20+, and
Python 3.11+ from their official or
distribution-supported sources. Install the Node tools:
npm install -g @anthropic-ai/claude-code playwright
npx playwright install --with-deps chromiumConfirm the active executables:
python3 --version # must be 3.11+
node --version # must be 20+
uv --version
ffmpeg -version
ffprobe -versionThe verification runner loads Playwright from the global npm root, so Chromium must be installed globally as shown above. Pinning Playwright as a versioned project dependency is tracked as a future improvement.
Clone and install GCA
git clone https://github.com/Tencent/Game-Creation-Agents.git
cd Game-Creation-Agents
cd gca-mcp
uv sync --extra dev
cd ..cp .env.example .envThe root .env file is loaded automatically by gca-mcp. Replace every
uncommented placeholder:
| Capability | Settings | Purpose |
|---|---|---|
| Image generation | OPENAI_API_KEY, GAME_ASSET_BASE_URL, GAME_ASSET_MODEL |
Generate and edit assets with the OpenAI Images API |
| Seedance video | ARK_API_KEY, SEEDANCE_URL, SEEDANCE_MODEL_STANDARD, SEEDANCE_MODEL_FAST |
Generate videos through Volcengine ARK |
| Video understanding | DOUBAO_API_ENDPOINT, DOUBAO_MODEL |
Route the pinned Doubao 2.0 model through ARK |
| Background removal (optional) | RMBG_BASE_URL, optional HF_TOKEN |
Not required: matting defaults to the model-free color_key route. Only used after you opt into assets.matting.backend: auto or rmbg2 โ see Background removal |
The OpenAI and ARK endpoints and model IDs in .env.example are provider
defaults, so most users only need to supply their API keys. The VLM policy
accepts only doubao-seed-2-0-pro-260215, with one request attempt per call.
DOUBAO_MODEL is pinned at runtime and is not user-configurable. Seedance
accepts local frames inline, so no object storage is required.
Store credential files securely and never commit .env or service-account JSON.
Verification is not optional-degradable: the VLM identity is pinned to Doubao
2.0 and there is no alternate provider chain. Without a working ARK_API_KEY,
screenshot and gameplay-video analysis fail rather than skip, which blocks the
verify phase of the pipeline. ARK keys are issued for the Volcengine
cn-beijing region; obtain one before running a full pipeline.
Restart the MCP server after changing .env, since it is read at startup.
GCA mattes with assets.matting.backend: color_key by default โ pure
computation against the generated #FF00FF / #00F000 staging background.
It needs no model, no API, and no credentials, and it downloads nothing. This
is the only route the project enables out of the box.
Color-key matting leaves harder edges than a segmentation model. For
soft-alpha edges you may opt into RMBG-2.0 by setting
assets.matting.backend to auto or rmbg2 in gca.config.yaml, and then
either pointing RMBG_BASE_URL at your own deployment or installing the
in-process extra:
cd gca-mcp && uv sync --extra local-rmbg --extra devImportant
RMBG-2.0 is licensed under CC BY-NC 4.0 and prohibits commercial use. GCA does not ship, bundle, or download it by default โ obtaining the model and complying with its license terms is your responsibility. See LICENSE.
Start Claude Code from the repository root:
claudeChoose either input style:
> Create a compact top-down arena game where a lighthouse keeper defends the beacon from shadow creatures. Use responsive controls and a short boss wave.Create game-brief.md in the repository root:
# Lighthouse Keeper
## Game
A compact top-down arena-defense game.
## Core loop
Protect the beacon from escalating shadow waves with responsive movement and
attacks. End the session with a short boss wave.
## Style
Moody coastal fantasy, readable combat, and a compact play session.Then ask Claude Code:
Create the game described in @game-brief.md.
Both options load GCA, run the required environment preflight, generate the game, and continue through verification automatically.
Each game is written to its own workspace:
playground/<game-slug>/
โโโ PROMPT.md # Original game request
โโโ GAME_DESIGN.md # Design and gameplay contracts
โโโ index.html # Browser entry point
โโโ package.json
โโโ src/ # Phaser scenes and game logic
โโโ assets/
โ โโโ asset-manifest.json # Runtime asset registry
โโโ levels/ # Present for authored level models
โโโ .playtest/
โโโ state.json # Current pipeline state
โโโ resume.json # Recovery information
โโโ tests/ # Test results
โโโ verify/ # Verification evidence and gates
Run the server from the repository root:
python3 scripts/serve_game.py playground/<game-slug> --base 3000The launcher stops this game's previous server, selects the first available port
starting at 3000, verifies that it is serving the requested workspace, and
prints the exact browser URL:
URL=http://127.0.0.1:<port>
Open that URL to play the generated game.
python3 scripts/serve_game.py playground/<game-slug> --stopThe preview server is for manual inspection. Canonical tests and release evidence are produced by the GCA verification pipeline.
GCA builds on the following core platforms:
| Project | Role in GCA |
|---|---|
| Claude Code | Loads the game-pipeline Skill and coordinates the specialized agents |
| Phaser | Runs each generated game in the browser; its Skills (MIT) are bundled as the pipeline's Phaser references |
| RMBG-2.0 | Optional, opt-in soft-alpha background removal (CC BY-NC 4.0; not bundled, not enabled by default) |
We thank their maintainers and contributors for the foundations that make this pipeline possible. We also thank OpenGame for serving as a design reference for GCA.
Copyright (C) 2026 Tencent. All rights reserved.
Game-Creation-Agents is licensed under the Apache License 2.0, except for the third-party components identified in LICENSE, which remain subject to their respective original license terms. This covers the repository source code, agent definitions, Skills, scripts, configuration, and documentation unless stated otherwise.
The bundled Phaser reference Skills under
.claude/skills/game-pipeline/references/phaser/ are derived from
Phaser and are licensed
under the MIT License โ see LICENSE for the full notice.
GCA ships no model weights. Background removal defaults to the model-free
color_key backend. RMBG-2.0 is an
optional third-party model that is never bundled, downloaded, or enabled by
default; it is licensed under CC BY-NC 4.0 and prohibits commercial use.
Users who opt into assets.matting.backend: auto or rmbg2 are responsible for
obtaining that model and complying with its license.
Third-party dependencies, supplied assets, and generated outputs remain subject to their respective licenses and usage terms.
If you find GCA helpful, please cite:
@article{gca2026,
title = {Game-Creation-Agents: Autonomous Game Production},
author = {Xiangxin Zhou and Jiaqi Tang and Ke Ma and Wenhao Yang and Jiarui Yao and Boye Niu and Haiyang Guo and Yifei Zhao and Huanjin Yao and Pu Jian and Ruihuang Li and Haitao Wu and Yushi Huang and Xiangyan Liu and Yan Sun and Kaixin Li and Tianyu Pang},
journal = {arXiv preprint},
year = {2026},
eprint = {2610.16776},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.16776}
}
