AI-Trader/CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

AI-Trader is an autonomous AI trading competition platform where multiple AI models compete in NASDAQ 100 trading with zero human intervention. Each AI starts with $10,000 and uses standardized MCP (Model Context Protocol) tools to make fully autonomous trading decisions.

Key Innovation: Historical replay architecture with anti-look-ahead controls ensures AI agents can only access data from the current simulation date and earlier.

Development Commands

Environment Setup

# Install dependencies
pip install -r requirements.txt

# Configure environment variables
cp .env.example .env
# Edit .env and set:
# - OPENAI_API_BASE, OPENAI_API_KEY
# - ALPHAADVANTAGE_API_KEY, JINA_API_KEY
# - AGENT_MAX_STEP (default: 30)

Data Preparation

# Download/update NASDAQ 100 stock data
cd data
python get_daily_price.py     # Fetch daily prices from Alpha Vantage
python merge_jsonl.py          # Merge into unified format (merged.jsonl)
cd ..

Starting Services

# Start all MCP services (Math, Search, Trade, LocalPrices)
cd agent_tools
python start_mcp_services.py
cd ..

# MCP services use fixed internal ports (8000-8003)
# These are not exposed to the host and should not be changed

Docker Deployment

# Build Docker image
docker-compose build

# Run with Docker Compose
docker-compose up

# Run in background
docker-compose up -d

# Run with custom config
docker-compose run ai-trader configs/my_config.json

# View logs
docker-compose logs -f

# Stop and remove containers
docker-compose down

# Pull pre-built image
docker pull ghcr.io/hkuds/ai-trader:latest

# Test local Docker build
docker build -t ai-trader-test .
docker run --env-file .env -v $(pwd)/data:/app/data ai-trader-test

Releasing Docker Images

# Create and push release tag
git tag v1.0.0
git push origin v1.0.0

# GitHub Actions automatically:
# 1. Builds Docker image
# 2. Tags with version and latest
# 3. Pushes to ghcr.io/hkuds/ai-trader

# Verify build in Actions tab
# https://github.com/HKUDS/AI-Trader/actions

Running Trading Simulations

# Run with default config
python main.py

# Run with custom config
python main.py configs/my_config.json

# Environment variables can override config dates:
INIT_DATE=2025-01-01 END_DATE=2025-01-31 python main.py

Complete Workflow

# All-in-one startup script (data + services + trading + web)
bash main.sh

Architecture

Core Components

1. Agent System (agent/base_agent/base_agent.py)

• BaseAgent: Base class for all trading agents
• Manages MCP tool connections, AI model initialization, trading execution loops
• Handles position management and logging
• Supports retry logic with exponential backoff (max_retries, base_delay)

2. Main Entry Point (main.py)

• Dynamic agent class loading via AGENT_REGISTRY
• Multi-model concurrent trading support
• Date range validation and weekday filtering
• Configuration management (JSON + environment variables)

3. MCP Toolchain (agent_tools/)

• tool_math.py: Mathematical calculations (port 8000)
• tool_jina_search.py: Market intelligence search (port 8001)
• tool_trade.py: Buy/sell execution (port 8002)
• tool_get_price_local.py: Price queries (port 8003)
• start_mcp_services.py: Service orchestration with health checks

4. Data Management (data/)

• daily_prices_*.json: Individual stock OHLCV data
• merged.jsonl: Unified price data format
• agent_data/[signature]/position/position.jsonl: Position records
• agent_data/[signature]/log/[date]/log.jsonl: Trading logs

5. Utilities (tools/)

• general_tools.py: Config management, message extraction
• price_tools.py: Price queries, position updates
• result_tools.py: Performance calculations

Data Flow

1. Initialization: Agent loads config, connects to MCP services, initializes AI model
2. Trading Loop: For each date:
   • Get system prompt with current positions, yesterday's prices, today's buy prices
   • AI agent analyzes market, calls search/math/price tools
   • Makes buy/sell decisions via trade tool
   • Logs all decisions and updates position.jsonl
3. Position Tracking: Each trade appends to position.jsonl with date, action, and updated holdings

Configuration System

Multi-layered config priority:

1. Environment variables (highest)
2. Model-specific config (openai_base_url, openai_api_key in model config)
3. JSON config file
4. Default values (lowest)

Runtime configuration (API mode only):

• Dynamically created per model-day execution via RuntimeConfigManager
• Isolated config files prevent concurrent execution conflicts
• Contains: TODAY_DATE, SIGNATURE, IF_TRADE, JOB_ID
• Written by write_config_value(), read by get_config_value()

Agent System

BaseAgent Key Methods:

• initialize(): Connect to MCP services, create AI model
• run_trading_session(date): Execute single day's trading with retry logic
• run_date_range(init_date, end_date): Process all weekdays in range
• get_trading_dates(): Resume from last date in position.jsonl
• register_agent(): Create initial position file with $10,000 cash

Adding Custom Agents:

1. Create new class inheriting from BaseAgent
2. Add to AGENT_REGISTRY in main.py:

   "CustomAgent": {
       "module": "agent.custom.custom_agent",
       "class": "CustomAgent"
   }
   

3. Set "agent_type": "CustomAgent" in config JSON

System Prompt Construction

Dynamic prompt generation (prompts/agent_prompt.py):

• get_agent_system_prompt() builds prompt with:
  • Current date
  • Yesterday's closing positions
  • Yesterday's closing prices
  • Today's buy prices
  • Yesterday's profit/loss
• AI agent must output <FINISH_SIGNAL> to end trading session

Anti-Look-Ahead Controls

Data access restrictions:

• Price data: Only returns data for date <= TODAY_DATE
• Search results: News filtered by publication date
• All tools enforce temporal boundaries via TODAY_DATE from runtime_env.json

Configuration File Format

{
  "agent_type": "BaseAgent",
  "date_range": {
    "init_date": "2025-01-01",
    "end_date": "2025-01-31"
  },
  "models": [
    {
      "name": "model-display-name",
      "basemodel": "provider/model-id",
      "signature": "unique-identifier",
      "enabled": true,
      "openai_base_url": "optional-override",
      "openai_api_key": "optional-override"
    }
  ],
  "agent_config": {
    "max_steps": 30,           // Max reasoning iterations per day
    "max_retries": 3,          // Retry attempts on failure
    "base_delay": 1.0,         // Base retry delay (seconds)
    "initial_cash": 10000.0
  },
  "log_config": {
    "log_path": "./data/agent_data"
  }
}

Data Formats

Position Record (position.jsonl):

{
  "date": "2025-01-20",
  "id": 1,
  "this_action": {
    "action": "buy",
    "symbol": "AAPL",
    "amount": 10
  },
  "positions": {
    "AAPL": 10,
    "MSFT": 0,
    "CASH": 9737.6
  }
}

Price Data (merged.jsonl):

{
  "Meta Data": {
    "2. Symbol": "AAPL",
    "3. Last Refreshed": "2025-01-20"
  },
  "Time Series (Daily)": {
    "2025-01-20": {
      "1. buy price": "255.8850",
      "2. high": "264.3750",
      "3. low": "255.6300",
      "4. sell price": "262.2400",
      "5. volume": "90483029"
    }
  }
}

Important Implementation Details

Trading Day Logic:

• Only weekdays (Monday-Friday) are processed
• get_trading_dates() automatically resumes from last date in position.jsonl
• Skips days already processed (idempotent)

Error Handling:

• All async operations use _ainvoke_with_retry() with exponential backoff
• MCP service failures raise detailed error messages with troubleshooting hints
• Missing API keys halt startup with clear error messages

Tool Message Extraction:

• extract_conversation(response, "final"): Get AI's final answer
• extract_tool_messages(response): Get all tool results
• Handles both dict and object-based message formats

Logging:

• Each trading day creates log/[date]/log.jsonl
• Logs include timestamps, signature, and all message exchanges
• Position updates append to single position/position.jsonl

Testing Changes

When modifying agent behavior or adding tools:

1. Create test config with short date range (2-3 days)
2. Set max_steps low (e.g., 10) to iterate faster
3. Check logs in data/agent_data/[signature]/log/[date]/
4. Verify position updates in position/position.jsonl
5. Use main.sh only for full end-to-end testing

Common Issues

MCP Services Not Running:

• Error: "Failed to initialize MCP client"
• Fix: cd agent_tools && python start_mcp_services.py
• Verify ports not already in use: lsof -i :8000-8003

Missing Price Data:

• Ensure data/merged.jsonl exists
• Run cd data && python get_daily_price.py && python merge_jsonl.py
• Check Alpha Vantage API key is valid

Runtime Config Issues:

• Runtime configs are automatically managed by the API
• Configs are created per model-day execution in data/ directory
• Ensure data/ directory is writable

Agent Doesn't Stop Trading:

• Agent must output <FINISH_SIGNAL> within max_steps
• Increase max_steps if agent needs more reasoning time
• Check log.jsonl for errors preventing completion