numb94/crewai-astock

CrewAI A-Stock is a research and learning project that uses a team of cooperating AI agents to analyze and recommend stocks listed on mainland China's A-share market. The README opens with a strong disclaimer: the project is for personal study and exploration of multi-agent patterns only, the AI outputs are not guaranteed to be accurate, and the code does not include any real trading interface, so it must not be used to move real money. The system is built on a framework called CrewAI, which lets you define a group of AI agents and have them cooperate on a task. This project defines seven roles: a review analyst that looks back at past recommendations, a market reporter that watches news and conditions, a stock picker that maintains a candidate pool, a multi-dimensional analyst covering technicals, capital flows, fundamentals, news, intraday data, and community sentiment, a risk manager with veto power, an investment decision agent that produces final scores and position sizes, and a position monitor that watches existing holdings and suggests sell points. These agents are grouped into three workflows: a recommendation crew, a single-stock evaluation crew, and a holdings monitoring crew. On top of the agents sits a web management panel built with Flask, Vue 3, Element Plus, and ECharts charts. It supports multiple users with thirty-day session isolation, shows real-time candlestick charts with recommendation markers, streams AI output through server-sent events, and tracks holdings, trades, and performance. Optional WeChat push notifications go through a service called PushPlus. A separate long-term memory feature uses BAAI bge-m3 embeddings on the SiliconFlow platform so that the crew can accumulate strategy lessons across runs. For the language model side the project uses DeepSeek for reasoning and optionally Grok for real-time news and social sentiment. Stock data comes from AKShare, a free open data library that needs no key, with an optional paid Zhitu API as a more stable alternative. News fallback uses Tavily. The storage layer is SQLite accessed through SQLAlchemy with fifteen tables. Getting it running needs Python 3.10 or newer. You clone the repo, set up a virtual environment, install requirements, add at least a DeepSeek key to the env file, initialize the database, optionally load basic info for around 5,161 A-share tickers, then start app.py on port 5000. The README also lists features removed from the open-source build, such as live broker integration with Tonghuashun or QMT and automatic buy and sell loops. The license is MIT.