Building an LLM Bot

The apps/llm-bot package demonstrates a poker bot powered by Claude AI that uses tool use for game analysis and opponent tracking. This guide covers the architecture and key concepts.

Features

• Claude-powered decisions — uses Claude Haiku for fast, intelligent poker decisions
• Tool use — Claude inspects game state, queries opponents, and calculates pot odds via tools
• Opponent tracking — maintains statistics on opponent tendencies (VPIP, PFR, aggression)
• Timeout fallback — graceful fallback to conservative play if Claude times out
• Reconnection — automatic reconnection on connection loss

Architecture

apps/llm-bot/src/
├── index.ts              # Entry point with environment config
├── bot.ts                # Main orchestrator
├── connection/           # WebSocket client
│   ├── ws-client.ts      # Connection handling
│   └── message-types.ts  # Protocol type definitions
├── agent/                # Claude integration
│   ├── llm-agent.ts      # Anthropic SDK wrapper
│   ├── model-factory.ts  # AI provider factory
│   ├── tools.ts          # Tool definitions for Claude
│   └── prompts.ts        # System prompts
├── game/                 # Game state tracking
│   ├── state-manager.ts  # State tracking
│   └── hand-utils.ts     # Card utilities
├── memory/               # Opponent profiling
│   ├── opponent-tracker.ts
│   └── types.ts
└── strategy/             # Fallback strategy
    └── fallback.ts

How It Works

1. Connection — Bot connects to Pokai server via WebSocket and registers
2. Match start — Receives match configuration and player list
3. Action required — When it's the bot's turn:
   • Claude receives the full game state and opponent profiles
   • Uses tools to analyze the situation
   • Submits action via the submit_action tool
4. Tracking — Records opponent actions for profiling
5. Round end — Updates opponent statistics with showdown data

Claude's Tools

The bot defines tools that Claude can call during decision-making:

This approach lets Claude reason about the game naturally — it can check the state, review opponent history, calculate odds, and then decide.

Opponent Profiling

The bot tracks detailed opponent statistics:

• VPIP (Voluntarily Put $ In Pot) — how often they play hands
• PFR (Pre-Flop Raise) — how often they raise pre-flop
• Aggression Factor — (Bets + Raises) / Calls
• Showdown hands — what cards they've shown down

From these stats, player styles are classified:

Fallback Strategy

If Claude times out (25s default), the bot uses a simple fallback:

1. CHECK if available (free action)
2. CALL if pot odds are favorable
3. FOLD otherwise

This prevents timeouts from costing the bot unnecessarily.

Quick Start

cd apps/llm-bot
cp .env.example .env
# Edit .env with your API keys
npm run dev

Environment Variables

Key Takeaways

• Tool use is powerful for poker bots — Claude can request exactly the information it needs
• Opponent tracking gives Claude context about player tendencies
• Fallback strategies prevent timeouts from being catastrophic
• Haiku is fast enough for real-time poker decisions (under 25s)