TradingAgents-MCPmode

# TradingAgents-MCPmode TradingAgents-MCPmode multi-agent trading analysis system based on MCP tools ## 🌟 Project Overview **TradingAgents-MCPmode** is an innovative multi-agent trading analysis system that integrates the Model Context Protocol (MCP) tool to achieve intelligent stock analysis and trading decision-making processes. The system provides comprehensive market analysis, investment advice, and risk management through the collaboration of 15 specialized intelligent agents. ### 🎯 Core Features - **🤖 Multi-Agent Collaboration**: 15 specialized intelligent agents collaborate - **⚡ Parallel Processing**: The analyst team adopts a parallel architecture, significantly improving analysis efficiency - **🔧 MCP Tool Integration**: Supports external data sources and real-time information acquisition - **📊 Comprehensive Analysis**: Seven-dimensional analysis of company overview, market, sentiment, news, fundamentals, shareholder structure, and product business - **💭 Intelligent Debate**: Bullish/bearish researcher debate mechanism, configurable debate rounds - **⚠️ Risk Management**: Three-layer risk analysis and management decision-making, supporting dynamic risk debate - **🎛️ Agent Control**: The front-end can dynamically enable/disable specific agents, flexibly customizing the workflow - **🌀 Debate Round Configuration**: The front-end can set investment and risk debate rounds in real-time, precisely controlling the depth of analysis - **🔧 Flexible Configuration**: Control agent MCP permissions through environment variables - **🌍 Multi-Market Support**: US stocks (US), A-shares (CN), Hong Kong stocks (HK) - **🗣️ Natural Language**: Supports natural language queries without specifying market and date - **📈 Real-Time Decision-Making**: Trading recommendations based on the latest data - **🌐 Web Front-End**: Complete Streamlit Web interface, supporting real-time analysis and historical management ## 🏗️ System Architecture ### Agent Organizational Structure ``` ┌─────────────────────────────────────────────────────────────┐ │ TradingAgents-MCPmode │ ├─────────────────────────────────────────────────────────────┤ │ 📊 Analysts - Parallel Execution │ │ ├── CompanyOverviewAnalyst │ │ ├── MarketAnalyst ┐ │ │ ├── SentimentAnalyst │ Parallel Processing │ │ ├── NewsAnalyst │ 6 Analysts │ │ ├── FundamentalsAnalyst│ Execute Simultaneously │ │ ├── ShareholderAnalyst │ │ │ └── ProductAnalyst ┘ │ ├─────────────────────────────────────────────────────────────┤ │ 🔬 Researchers │ │ ├── BullResearcher │ │ └── BearResearcher │ ├─────────────────────────────────────────────────────────────┤ │ 👔 Managers │ │ ├── ResearchManager │ │ └── Trader │ ├─────────────────────────────────────────────────────────────┤ │ ⚠️ Risk Management │ │ ├── AggressiveRiskAnalyst │ │ ├── SafeRiskAnalyst │ │ ├── NeutralRiskAnalyst │ │ └── RiskManager │ └─────────────────────────────────────────────────────────────┘ ``` ### 🚀 Parallelized Workflow ```mermaid graph TD %% 用户输入 USER[👤 用户输入<br/>user_query] %% 第0阶段：公司概述分析师 A0[🏢 公司概述分析师<br/>📥 输入：user_query<br/>📤 输出：company_details + company_overview_report] %% 第1阶段：分析师并行节点 PARALLEL[⚡ 分析师并行节点<br/>📥 输入：user_query + company_details<br/>🔄 并发执行6个分析师<br/>📤 输出：全部6个分析报告] %% 并行执行的6个分析师 A1[🔍 市场分析师] A2[😊 情绪分析师] A3[📰 新闻分析师] A4[📊 基本面分析师] A5[👥 股东分析师] A6[🏭 产品分析师] %% 第2阶段：研究员辩论 B1[📈 看涨研究员<br/>📥 输入：user_query + 全部7个分析师报告<br/>📤 输出：看涨论证 + 辩论历史] B2[📉 看跌研究员<br/>📥 输入：user_query + 全部7个分析师报告 + 辩论历史<br/>📤 输出：看跌论证 + 辩论历史] %% 第3阶段：管理层 C1[🎯 研究经理<br/>📥 输入：user_query + 全部7个分析师报告 + 完整辩论历史<br/>📤 输出：investment_plan] C2[💰 交易员<br/>📥 输入：user_query + 全部7个分析师报告 + 辩论历史 + investment_plan<br/>📤 输出：trader_investment_plan] %% 第4阶段：风险管理团队 D1[🔥 激进风险分析师<br/>📥 输入：全部信息<br/>📤 输出：激进风险观点 + 风险辩论历史] D2[🛡️ 保守风险分析师<br/>📥 输入：全部信息 + 风险辩论历史<br/>📤 输出：保守风险观点 + 风险辩论历史] D3[⚖️ 中性风险分析师<br/>📥 输入：全部信息 + 风险辩论历史<br/>📤 输出：中性风险观点 + 风险辩论历史] D4[🎯 风险经理<br/>📥 输入：全部信息 + 完整风险辩论历史<br/>📤 输出：final_trade_decision] %% 主要流程连接 USER --> A0 A0 --> PARALLEL PARALLEL --> B1 B1 --> B2 B2 -.->|辩论循环| B1 B1 --> C1 C1 --> C2 C2 --> D1 D1 -.->|风险辩论循环| D2 D2 -.->|风险辩论循环| D3 D3 -.->|风险辩论循环| D1 D1 --> D4 D2 --> D4 D3 --> D4 %% 并行节点内部连接（虚线表示并发） PARALLEL -.-> A1 PARALLEL -.-> A2 PARALLEL -.-> A3 PARALLEL -.-> A4 PARALLEL -.-> A5 PARALLEL -.-> A6 %% 样式定义 style USER fill:#f9f9f9,stroke:#333,stroke-width:2px style A0 fill:#e1f5fe,stroke:#0277bd style PARALLEL fill:#fff3e0,stroke:#ff8f00,stroke-width:3px style A1 fill:#e8f5e8,stroke:#4caf50 style A2 fill:#e8f5e8,stroke:#4caf50 style A3 fill:#e8f5e8,stroke:#4caf50 style A4 fill:#e8f5e8,stroke:#4caf50 style A5 fill:#e8f5e8,stroke:#4caf50 style A6 fill:#e8f5e8,stroke:#4caf50 style B1 fill:#e3f2fd,stroke:#1976d2 style B2 fill:#e3f2fd,stroke:#1976d2 style C1 fill:#fff3e0,stroke:#ef6c00 style C2 fill:#fff3e0,stroke:#ef6c00 style D1 fill:#fce4ec,stroke:#c2185b style D2 fill:#fce4ec,stroke:#c2185b style D3 fill:#fce4ec,stroke:#c2185b style D4 fill:#fce4ec,stroke:#c2185b ``` ### ⚡ Parallelization Advantages 1. **Significantly Improved Efficiency**: 6 analysts execute concurrently, and the total time is close to the execution time of the slowest analyst 2. **Resource Optimization**: Fully utilize system resources and avoid serial waiting 3. **Data Consistency**: Each analyst uses a deep copy state to avoid concurrent conflicts 4. **Result Merging**: Intelligently merge the reports and execution history of each analyst #### Online Experience (Public Network Address) - Experience Address: `http://47.79.147.241:8501` - Description: Open to the public for free for function experience and demonstration. - Note: Please use it civilly and do not conduct any form of attack, stress test, or malicious access behavior to avoid affecting service stability. <img width="1920" height="1080" alt="image" src="https://github.com/user-attachments/assets/c557e15a-de1e-4dbe-b619-b8f32a2a91c2" /> ## 🚀 Quick Start ### Environment Requirements - Python 3.8+ - Supported operating systems: Windows, macOS, Linux ### Installation Steps 1. **Clone the Project** ```bash git clone https://github.com/guangxiangdebizi/TradingAgents-MCPmode.git cd TradingAgents-MCPmode ``` 2. **Install Dependencies** ```bash pip install -r requirements.txt ``` 3. **Configure Environment Variables** ```bash cp env.example .env # Edit the .env file to configure your API keys and workflow parameters ``` 4. **Configure MCP Tool** ```bash # Edit the mcp_config.json file to configure the MCP server ``` ### 🌐 Web Front-End Usage **Recommended Usage** - Interact through the Web interface: ```bash streamlit run web_app.py ``` Then visit in your browser: `http://localhost:8501` <img width="1920" height="1080" alt="image" src="https://github.com/user-attachments/assets/280599ee-c00a-4c8c-b61f-c786e43cc6d4" /> #### Web Front-End Features - **🏠 Home**: System overview and quick navigation - **⚙️ System Configuration**: Online editing of .env and MCP configuration - **🔍 Real-Time Analysis**: Enter queries and monitor analysis progress in real-time - **🤖 Agent Control**: Dynamically enable/disable specific agents, checkbox interface organized by team - **🌀 Debate Configuration**: Real-time adjustment of investment and risk debate rounds, precisely controlling the depth of analysis - **📊 Agent Page**: View detailed analysis results of each agent by team - **📚 Historical Reports**: Manage historical sessions, support Markdown, PDF, DOCX export #### 🎛️ Front-End Agent Control Features **Agent Enablement Control**: - ✅ **Group Management**: Organized by four teams: analysts, researchers, management, and risk management - ✅ **Real-Time Selection**: Checkbox interface, supporting quick operations for selecting all/none - ✅ **Dynamic Counting**: Real-time display of the number of enabled agents (e.g., Enabled 12/15) - ✅ **Intelligent Skipping**: Disabled agents are automatically skipped, without affecting workflow continuity **Debate Round Configuration**: - 🔄 **Investment Debate**: Slider settings for bullish/bearish researcher debate rounds (each round = each side speaks once) - ⚖️ **Risk Debate**: Slider settings for aggressive/conservative/neutral risk analyst debate rounds (each round = each side speaks once) - 📊 **Real-Time Preview**: Display current settings and expected number of speeches - 🚀 **Automatic Application**: Automatically apply front-end settings when starting analysis, without manual confirmation ### 💻 Command Line Usage If you prefer the command line: ```bash python main.py -c "Analyze Apple's stock" ``` #### Interactive Mode ```bash python main.py ``` Then enter your query as prompted, for example: - "Analyze Tesla stock" - "Analyze 600036 China Merchants Bank for me" - "Analyze the investment value of NVDA" ## 📊 Data Flow Mechanism ### 🎯 Key Design Features 1. **company_details Placeholder System** - 🏢 **Company Overview Analyst** first obtains basic company information - 📥 **Only passed to the analyst team**: 6 professional analysts can obtain accurate company background - 🚫 **Not passed to subsequent agents**: Researchers and subsequent agents focus on comprehensive evaluation of analysis reports 2. **Information Accumulation Effect** - **Stage 0**: User query only - **Stage 1**: User query + company details (parallel processing) - **Stage 2**: User query + all 7 analyst reports - **Stage 3**: User query + analyst reports + debate history + investment decisions - **Stage 4**: User query + all information + risk opinions 3. **Dual Debate Mechanism** - **Investment Debate**: Bullish ↔ Bearish researchers debate in a loop - **Risk Debate**: Aggressive ↔ Conservative ↔ Neutral risk analysts debate in a loop ### 💡 Practical Application Example **User Input**: "Analyze Tesla stock" 1. **🏢 Company Overview Analyst**: Obtain basic information such as Tesla Inc., TSLA, NASDAQ, and the electric vehicle industry 2. **⚡ Analyst Parallel Node**: 6 analysts simultaneously conduct professional analysis based on accurate "Tesla Inc." information 3. **📈 Bullish Researcher**: Synthesize all 7 professional analysis reports to construct a bullish investment argument for Tesla 4. **🎯 Risk Manager**: Make the final risk decision on Tesla investment based on all information and debate results This design ensures **precise information transmission** and **gradual refinement**, allowing each agent to exert professional capabilities in the most suitable information environment! ## 🎯 Usage Examples ### 📱 Front-End Operation Example **Scenario**: Quickly analyze Apple's stock, enable only core analysts, and set up simplified debates 1. **Start the System**: ```bash streamlit run web_app.py ``` 2. **Configure Agents** (expand in "🤖 Agents Enabled This Round"): - ✅ Analyst Team: Select All (7 agents) - ✅ Researcher Team: Select Bullish Researcher only - ❌ Management: Select None (quick analysis) - ❌ Risk Management: Select None (quick analysis) - 📊 Display: Enabled 8/15 3. **Set Debate Rounds** (expand in "🌀 Debate Round Settings"): - 🔄 Investment Debate Rounds: 1 round (only the bullish researcher speaks once) - ⚖️ Risk Debate Rounds: 0 rounds (skip risk debate) 4. **Execute Analysis**: - Input: "Analyze the investment value of Apple's stock" - Click "🚀 Start Analysis" - The system automatically applies the front-end settings and starts execution 5. **Expected Results**: - ⚡ Quick Execution: Only 8 agents participate, significantly reducing analysis time - 📊 Core Reports: Obtain 7 professional analyst reports - 🎯 Investment Advice: The bullish researcher provides investment advice based on analyst reports ### 🎛️ Dynamic Configuration Advantages - **🚀 Quick Experience**: Disable non-core agents to get analysis results in 1-2 minutes - **🔬 In-Depth Analysis**: Enable all agents + multiple rounds of debate to get complete decision support - **🎯 Customization**: Flexibly select agent combinations based on analysis needs - **💡 Real-Time Adjustment**: No need to restart the service, adjust parameters in real-time on the front-end ## 🛠️ Configuration Instructions ### Environment Variable Configuration Configure the following parameters in the `.env` file: ```env # Large Model Configuration OPENAI_API_KEY=your_openai_api_key OPENAI_BASE_URL=https://api.openai.com/v1 MODEL_NAME=gpt-4 # Workflow Configuration MAX_DEBATE_ROUNDS=1 # Default number of investment debate rounds (can be dynamically adjusted on the front-end) MAX_RISK_DEBATE_ROUNDS=1 # Default number of risk debate rounds (can be dynamically adjusted on the front-end) DEBUG_MODE=true # Debug mode, display detailed logs VERBOSE_LOGGING=true # Detailed log output # Agent MCP Permission Configuration # Analyst Team - It is recommended to enable MCP to obtain real-time data COMPANY_OVERVIEW_ANALYST_MCP=true MARKET_ANALYST_MCP=true SENTIMENT_ANALYST_MCP=true NEWS_ANALYST_MCP=true FUNDAMENTALS_ANALYST_MCP=true SHAREHOLDER_ANALYST_MCP=true PRODUCT_ANALYST_MCP=true # Researcher/Management/Risk Team - It is recommended to turn off MCP to focus on comprehensive analysis BULL_RESEARCHER_MCP=false BEAR_RESEARCHER_MCP=false RESEARCH_MANAGER_MCP=false TRADER_MCP=false AGGRESSIVE_RISK_ANALYST_MCP=false SAFE_RISK_ANALYST_MCP=false NEUTRAL_RISK_ANALYST_MCP=false RISK_MANAGER_MCP=false # Task Quantity Concurrency Limit MAX_CONCURRENT_ANALYSIS=2 # Number of analysis tasks running simultaneously ``` > **💡 Tip**: The `MAX_DEBATE_ROUNDS` and `MAX_RISK_DEBATE_ROUNDS` in the environment variables are only default values. The actual runtime can be adjusted in real-time on the front-end. It is recommended to set it to 1 for a quick experience. If you need in-depth analysis, you can increase the number of rounds on the front-end. ### MCP Tool Configuration Configure the MCP server in the `mcp_config.json` file: ```json { "mcpServers": { "finance-mcp": { "disabled": false, "timeout": 600, "transport": "streamable_http", "url": "https://finvestai.top/mcp", "headers": { "X-Tushare-Token": "Your tushare token" } } } } ``` **Configuration Instructions**: - `disabled`: Whether to disable the MCP server (false means enabled) - `timeout`: Request timeout time (seconds), default is 600 seconds - `transport`: Transport protocol, use `streamable_http` - `url`: MCP server address `https://finvestai.top/mcp` - `X-Tushare-Token`: Please replace with your real Tushare API Key > **💡 Tip**: After modifying the configuration, please restart the application or hot reload the MCP client to make the configuration take effect. ## 📈 Performance Optimization ### Parallel Processing Advantages - **Original Serial Architecture**: 6 analysts execute sequentially, and the total time = Σ(time of each analyst) - **New Parallel Architecture**: 6 analysts execute concurrently, and the total time ≈ max(time of each analyst) - **Performance Improvement**: Theoretically, the efficiency can be improved by 5-6 times, and the actual improvement depends on the time distribution of each analyst ### Resource Management - Use deep copies to avoid state competition - Intelligently merge execution history and tool call records - Optimize memory usage and avoid state redundancy ## 🤝 Contribution Guide Welcome to submit Issues and Pull Requests to improve the project! ## 📄 License This project is licensed under the MIT License. See the [LICENSE](LICENSE) file for details. --- **TradingAgents-MCPmode** - Let the AI agent team escort your investment decisions! 🚀📈 ## 📬 Contact Information - Email: `guangxiangdebizi@gmail.com` - LinkedIn: [Xingyu Chen](https://www.linkedin.com/in/xingyu-chen-b5b3b0313/) - Facebook: [Personal Homepage](https://www.facebook.com/profile.php?id=100072282093932) - Bilibili: [Personal Space](https://space.bilibili.com/51239486?spm_id_from=333.1007.0.0)