LMChess Arena

Challenge Beyond Human thinking and Intelligence

Comment

Inspiration

We were inspired by the idea of transforming chess from a silent, calculation-heavy game into an interactive, personality-driven experience. Traditional chess engines like Stockfish are incredibly strong but completely emotionless. We wanted to create something different—where your opponent isn't just calculating moves, but trash-talking, strategizing aloud, and reacting dynamically to your gameplay. By leveraging Google's Gemini AI and other leading language models, we reimagined chess as a conversational battle where every move comes with commentary, creating an engaging experience that feels like playing against a real grandmaster with personality.

What it does

LMC Arena is an AI-powered chess platform where players battle against leading language models (Gemini, GPT, NVIDIA) that act as chess opponents with distinct personalities. The platform:

  • Enforces all legal chess moves with real-time visual feedback (highlighted squares, blocked illegal moves)
  • Provides AI commentary after each move, making the opponent feel alive and engaging
  • Features AutoPilot mode where two AIs battle each other while you spectate
  • Tracks comprehensive stats including wins, losses, draws, and match history
  • Implements daily play limits (3 free matches per day) to manage API costs responsibly
  • Offers draw mechanics where players can propose draws after 10 moves, and the AI can accept or decline
  • Displays captured pieces and move history in a clean, Chess.com-inspired interface
  • Supports multiple AI models with different playing styles and personalities

How we built it

Frontend:

  • Built with React 18 for modern UI components and state management
  • Used chess.js library for complete chess rule validation and legal move generation
  • Styled with Tailwind CSS 3 for a responsive, neo-brutalist design
  • Implemented Firebase SDK for authentication (email/password + Google OAuth)
  • Created custom components for board, squares, move history, and player info

Backend:

  • Set up Firebase Cloud Functions (Node.js) as serverless API endpoints
  • Integrated Google Gemini 2.0 Flash API for primary AI opponent
  • Connected OpenRouter API to access GPT-4o-mini and NVIDIA models
  • Implemented Cloud Firestore for storing users, matches, and daily limits
  • Created security rules to protect user data

AI Integration:

  • Crafted persona-driven prompts for each AI model (arrogant grandmaster, tactical warrior, etc.)
  • Built validation system to catch AI hallucinations (invalid moves) with fallback to random legal moves
  • Implemented retry logic and error handling for API failures
  • Added commentary extraction to display AI's thoughts after each move

Deployment:

  • Deployed frontend to Vercel for global CDN distribution
  • Used Firebase Hosting as backup deployment option
  • Set up environment variables for secure API key management
  • Configured CORS and domain authorization

Challenges we ran into

  1. AI Hallucinations: Language models sometimes suggested illegal chess moves (e.g., moving knights like bishops). We solved this by validating every AI move with chess.js and implementing a fallback system that plays a random legal move when the AI makes an error.

  2. API Rate Limits: Gemini and OpenRouter APIs have usage quotas. We implemented a daily match limit (3 games per day) and optimized API calls to reduce costs while maintaining a good user experience.

  3. Real-time Move Validation: Ensuring users could only make legal moves required complex state management. We built a system that highlights legal moves when a piece is selected and blocks invalid moves with helpful toast notifications.

  4. Firebase Cold Starts: Cloud Functions had initial latency issues. We optimized by keeping functions warm and implementing loading states with "AI is thinking..." indicators.

  5. Prompt Engineering: Getting consistent JSON responses from AI models was tricky. We refined prompts to explicitly request JSON-only output and added parsing logic to strip markdown formatting.

  6. Draw Mechanics: Implementing the 50-move rule was costly (50 API calls without progress). We reduced it to 15 moves for Gemini to balance gameplay and API costs.

  7. State Synchronization: Managing board state, move history, captured pieces, and AI commentary required careful React state management to avoid bugs and race conditions.

Accomplishments that we're proud of

  • 100% Legal Move Enforcement: Every single move is validated—no illegal moves possible
  • Multi-Model Support: Successfully integrated 3 different AI providers (Google, OpenAI via OpenRouter, NVIDIA)
  • AutoPilot Mode: A unique feature where two AIs battle each other, showcasing creative AI vs AI gameplay
  • Professional UI: Built a Chess.com-quality interface with smooth animations and intuitive controls
  • Robust Error Handling: Graceful fallbacks for API failures, invalid moves, and network issues
  • Complete Authentication Flow: Secure login with email/password and Google OAuth
  • Comprehensive Stats Tracking: Full game history with wins/losses/draws stored in Firestore
  • Draw Negotiation System: First chess app where you can propose draws and AI responds dynamically
  • Scalable Architecture: Firebase backend scales automatically with user growth
  • Daily Limit System: Fair play mechanism that balances free usage with API costs

What we learned

  • Prompt Engineering is Critical: Small changes in AI prompts drastically affect move quality and commentary style
  • Firebase is Powerful: Cloud Functions + Firestore + Auth = complete backend with minimal setup
  • Chess.js is Essential: Would be nearly impossible to implement legal moves without this library
  • AI Models Have Different Personalities: Gemini plays aggressive, GPT plays tactical, NVIDIA plays analytical—even with similar prompts
  • User Feedback Matters: Toast notifications and visual highlights dramatically improve user experience
  • Rate Limiting is Necessary: Without daily limits, API costs would be unsustainable for a free app
  • State Management is Complex: Chess requires careful tracking of board state, turn order, captured pieces, and game status
  • Deployment Best Practices: Environment variables, CORS configuration, and security rules are crucial for production apps
  • Error Handling Saves Projects: AI APIs fail sometimes—fallback logic keeps the app running smoothly

What's next for LMChess Arena

Short-term (Next 1-2 Months):

  • Add more AI models: Claude 3.5, LLaMA 3, Mistral, DeepSeek
  • Implement ELO ratings: Track skill levels for both users and AI models
  • Add difficulty levels: Adjust AI temperature and thinking time for easier/harder opponents
  • Voice commands: Use Google Speech-to-Text for move input ("Knight to F3")
  • Mobile app: React Native version for iOS/Android
  • Leaderboard: Global rankings for players with most wins

Medium-term (3-6 Months):

  • Multiplayer mode: Play against other humans with AI commentary
  • Tournament system: Weekly tournaments with prizes
  • Game analysis: Post-game breakdown showing best moves and mistakes
  • Opening library: Learn chess openings with AI coaching
  • Puzzle mode: Daily chess puzzles with AI hints
  • Payment integration: Subscription for unlimited matches

Long-term (6-12 Months):

  • Computer vision: Scan physical chess boards with phone camera
  • AR mode: Overlay AI suggestions on real chess boards using AR
  • Live streaming: Watch top players battle AI opponents
  • AI training: Upload your games to create a personalized AI that plays like you
  • Custom AI personas: Create your own trash-talking AI opponents
  • Chess variants: Support for Chess960, 3-check, King of the Hill
  • Social features: Friends, chat, game sharing, replays

Built With

Share this project:

Updates