AI Game Chest
A local-first AI games collection: 50+ browser games, AI opponents with real engines, and 107 OpenSpiel games. All running on your machine — no accounts, no ads, no cloud.
Chess with Stockfish 16 (3500+ Elo) · Go with KataGo · Shogi with YaneuraOu · Othello with Edax 4.6
Backgammon with GNU Backgammon · Hex with MoHex · 119 games via OpenSpiel
Japanese: kanji, JLPT N5-N1, flashcards, vocabulary · 23 board games, 19 arcade, 10+ card games
Current Development — Universal Game Replay & Analysis
Every nontrivial game is getting move-by-move replay, auto-save to localStorage, resume after page reload, and AI analysis (per-move: what would the engine have played?).
| Phase | Games | Status |
|---|
| 1 — Canary | Chess, Hex | ✅ Done |
| 2 — Core | Checkers, Othello, Go | ✅ Done |
| 3 — Fleet | All 2-player games | 🔜 2/17 wired (connect4, shogi) |
| 4 — AI Analysis | Engine comparison per move | 📋 Planned |
| 5 — UI | Game library page, export, comments | 📋 Planned |
See docs/plan/game-replay-analysis-plan.md.
Why This Exists
This started as an experiment: how far can agentic engineering go? Not "vibecoding" — deliberate, iterative, push-the-limits AI-assisted development. The result is a full-featured games platform with 150+ games, seven AI engines, 3D physics, and an MCP API. All built in days, not months. This repo is the artifact of that question.
The AI opponent irony. Every AI engine here — Stockfish, KataGo, Edax, MoHex, GNU Backgammon — was built by human researchers over decades. Not a single line of their game-playing code was written by AI. They're classical algorithms: alpha-beta pruning, Monte Carlo tree search, neural network evaluation. Yet they are artificial intelligence in the truest sense. This project respects that enormous human effort and wraps it in a modern interface so it doesn't rot in a terminal.
Agent-accessible. Every game is also an MCP tool. Claude, Cursor, or any MCP client can analyze chess positions, play a round of Hex, or run a Go analysis. The /mcp endpoint exposes the entire collection programmatically. Games become building blocks for autonomous agents.
Local first. Everything runs on your machine. No account creation, no data collection, no ads, no monthly subscription. Start the server, open the browser, play.
107+ game research library. OpenSpiel is an academic framework from Google DeepMind. This is the only browser UI for it — you can play Phantom Tic-Tac-Toe, Liar's Dice, Breakthrough, and 100+ other games that exist nowhere else as clickable web pages.
3D physics for the sake of it. Jenga with Cannon.js physics. Mahjong with Three.js orbit controls. The kind of thing you build when you're not optimizing for mobile battery life.
SOTA Components
This repo doubles as a working portfolio of current-generation engineering practices:
| Layer | Technology | What it does |
|---|
| Web framework | React 19 + Vite 7 + TypeScript 5.9 | Modern SPA, instant HMR |
| 3D rendering | Three.js r128 + Cannon.js | Jenga physics, Mahjong orbit controls |
| Canvas | Custom 2D canvas | Hex board, Jigsaw puzzle pieces |
| Backend | FastAPI + FastMCP 3.2 | REST API + MCP dual transport |
| AI/ML | Stockfish 16, KataGo, MCTS, Edax, MoHex, GNU Bg | Classical game AI (not LLM) |
| Academic | OpenSpiel 1.6 (Google DeepMind) | 107 game environments with MCTS |
| Desktop | Tauri 2.0 + NSIS | Single-installer Windows app |
| Build | PyInstaller (frozen backend) | Standalone .exe, no Python runtime |
| Container | Docker Compose | Full stack engine deployment |
| Package | uv (Astral) | Lockfile-venv Python management |
| CI | GitHub Actions | Lint, typecheck, test, build |
| Task runner | Justfile | 15+ recipes for dev/build/test |
| E2E | Playwright | Automated browser tests |
| MCP | Model Context Protocol | Agent-accessible game API |
| PWA | Service worker + manifest | Offline-capable, installable |
| Multiplayer | WebSocket + Firebase | Real-time P2P game sessions |
| Quality | Ruff, tsc --noEmit, Bandit | Lint, typecheck, security audit |
| Git | Conventional commits | Structured changelog generation |
| Styling | CSS custom properties + dark theme | Consistent dark UI, no framework |
Table of Contents
Preview
50+ games across 12 categories.
| |
|---|
| |
| 3D Chess with Three.js | Star Trek TDC |
| |
| Japanese flower cards | Austrian card classic |
Features
- Play against real AI — Stockfish 16 (chess), KataGo (Go), YaneuraOu (shogi), Edax 4.6 (Othello), GNU Backgammon, MoHex (Hex), OpenSpiel (119 games). No JavaScript toy engines.
- 50+ browser games — chess, Go, shogi, poker, mahjong, arcade, puzzles, card games
- Learn Japanese — 2,500 kanji, JLPT practice (N5-N1), spaced-repetition flashcards
- MCP tools for agents — 15 FastMCP 3.2 tools for game analysis, coaching, tournaments, shared sessions
- Desktop app — Tauri 2.0, single NSIS installer, everything embedded
- Docker — all engines + gateway in containers
- P2P multiplayer — Firebase RTDB-synced sessions (
games/{id} nodes) shared between the browser games, MCP tools (list_shared_sessions, join_shared_session) and the dashboard lobby. Status + next steps: docs/MULTIPLAYER_STATUS.md
Game AI Engines
All engines run in Docker containers (orchestrated via docker-compose.yml) or natively on Windows. Each engine has a Python aiohttp server wrapper exposing a REST API on its port.
| Engine | Game | Version | License | Port | Windows | macOS | Docker Required? |
|---|
| Stockfish | Chess | 16 | GPL-3.0 | 11211 | ✅ Built-in | ✅ brew install stockfish | No |
| YaneuraOu | Shogi | 9.40 | GPL-3.0 | 11212 | ✅ Built-in | ✅ Built-in | No |
| KataGo | Go | 1.16.5 | MIT | 11213 | ✅ Built-in | ✅ brew install katago | No |
| Edax | Othello | 4.6 | GPL-3.0 | 11214 | ✅ engines/data/ | ❌ | Optional |
| OpenSpiel | 119 games | 1.6.15 | Apache-2.0 | 11216 | ✅ pip | ✅ pip | No |
| GNU Backgammon | Backgammon | 1.08 | GPL-3.0 | 11215 | ❌ | ✅ brew install gnubg | Optional (Linux-only) |
| MoHex | Hex | Source | LGPL-3.0 | 11210 | ❌ | ✅ scripts/build-mohex-macos.sh | Yes (or build native) |
Start all engines:
docker compose up -d
Naked-PC alternative (run engine servers directly without Docker):
uv run python engines/stockfish-server.py
uv run python engines/shogi-server.py
uv run python engines/go-server.py
uv run python engines/edax-server.py
uv run python engines/gnubg-server.py
uv run python engines/open_spiel_server.py
uv run python engines/mohex-server.py
Quick Start
git clone https://github.com/sandraschi/ai-game-chest
cd ai-game-chest
.\start.ps1
This launches all seven AI engines (including OpenSpiel on 11216) + the game gateway. Opens http://localhost:10987/ — browse 180+ games, play chess vs Stockfish, shogi vs YaneuraOu, or any of 107 OpenSpiel environments. Board-game multiplayer runs over Firebase (anonymous sign-in; enable it in the Firebase console → Authentication → Anonymous).
For other install methods (manual, Docker, Tauri desktop), see INSTALL.md.
What You Can Do
Play a game of chess against Stockfish at level 15
Study kanji stroke order for JLPT N3 vocabulary
Challenge KataGo to a 9x9 Go game
Browse the arcade — 19 classic games from Snake to Pac-Man
Open the MCP dashboard at http://localhost:10986
Game Collection
Arcade, board, card, casino, puzzle, strategy, and multiplayer games — all in your browser. Browse the collection.
Japanese Learning
Kanji wall (2,500), JLPT N5-N1 test drills, vocabulary with spaced repetition, stroke order, grammar, and listening practice. Start learning.
MCP Server
FastMCP 3.2 server with 15 tools for AI-assisted game analysis, coaching, tournaments, and P2P multiplayer. Server docs.
Documentation
Requirements
| Dependency | Install |
|---|
| Python 3.13+ | winget install Python.Python.3.13 |
| uv | winget install astral-sh.uv |
| Node.js 20+ | winget install OpenJS.NodeJS.LTS |
| just | winget install Casey.Just |
References
A full bibliography covering the AI engines, game theory, neurobiology of decision-making, game philosophy, and technical foundations is in REFERENCES.md. Highlights:
- OpenSpiel: Lanctot et al., "A Framework for Reinforcement Learning in Games" — arXiv:1908.09453
- Stockfish: Romstad et al., open-source chess engine — stockfishchess.org
- KataGo: Wu, "A Distributed Training Approach for Computer Go" — GitHub
- MCTS: Browne et al., "A Survey of Monte Carlo Tree Search Methods" — IEEE CIG
- Game philosophy: Huizinga, Homo Ludens (1938); Suits, The Grasshopper (1978); Caillois, Man, Play and Games (1961)
- Game theory: von Neumann & Morgenstern (1944); Nash (1950)
- Decision neuroscience: Kahneman, Thinking, Fast and Slow (2011); Damasio, Descartes' Error (1994)
License
MIT
