multiagent-stock-research

Multi-agent equity research workflow for Claude Code — inspired by TradingAgents, rebuilt as a buy-side analyst command with a Goldman-Sachs-style HTML report.

English | 中文

---

English

What it is

A single /analyze NVDA command in Claude Code produces an institutional-grade research report with:

• 13 analysis modules — macro environment, fundamentals, management quality, earnings momentum, valuation (DCF / comps / SOTP / reverse-DCF), industry positioning, catalysts, technical levels + institutional flow, risk sizing, balance-sheet health, cross-asset correlations, bubble detection, sentiment
• Module 10: Multi-agent Bull vs Bear debate (3 rounds: opening, rebuttal, closing) judged by a portfolio-manager agent
• Self post-mortem — reads your past reports on the same ticker, judges which predictions came true, feeds that bias-correction into the debate
• Goldman-Sachs-style HTML report — 21 sections, Goldman color palette, print-ready A4
• Markets: US / HK / A-shares

vs TradingAgents

	multiagent-stock-research	TradingAgents
Analysis depth	13 modules	4 analysts
Debate rounds	3 (open / rebuttal / closing)	configurable, default 1
Markets	US + HK + A-shares	US only
Self-reflection	yes — reads own past reports	no
Output format	Goldman-style HTML + scorecard + stop-loss triggers	trade decision string
Methodology frameworks	Graham / Buffett / O'Neil / Minervini / Druckenmiller / Marks	generic LLM prompting
Install	1 shell script, no Python setup	Python package + Docker
Runtime deps	none (Claude Code only)	LangGraph + multiple LLM SDKs

Quick start

# 1. Clone
git clone https://github.com/YOUR_USER/multiagent-stock-research
cd multiagent-stock-research

# 2. Install
./install.sh

# 3. Configure (add to ~/.zshrc or ~/.bashrc)
export TUSHARE_TOKEN="..."      # A-share analyst data — https://tushare.pro
export FINNHUB_TOKEN="..."      # US insider/institutional — https://finnhub.io
export RESEARCH_OUTPUT_DIR="$HOME/equity-research"   # optional, default ~/equity-research

# 4. Use in Claude Code
/analyze NVDA --html

Commands

Command	Behavior
/analyze TICKER	Full 13 modules + debate, markdown output
/analyze TICKER --html	Also generate HTML report
/analyze TICKER --silent	No console output, progress lines only
/analyze TICKER --quick	Skip Module 10 debate (fast screening)
/analyze TICKER --lang en	Output the report in English (default: 中文)
/analyze T1,T2,T3 --html --silent	Batch mode
/report TICKER	Regenerate HTML from latest analysis

Sample output

A real rendered report is at samples/AAPL-sample-report.html. Open it in a browser to see the full Goldman-Sachs-style layout, including Module 10 debate (Bull/Bear columns, core disagreement matrix, unresolved risks, composite scorecard with debate adjustments).

Quick preview without cloning: https://raw.githack.com/chenhab03/multiagent-stock-research/main/samples/AAPL-sample-report.html

English sample report (generated with `--lang en`): samples/AAPL-sample-english.html — preview: https://raw.githack.com/chenhab03/multiagent-stock-research/main/samples/AAPL-sample-english.html

Architecture

See docs/debate-mechanism.md for the debate flow and docs/methodology.md for the 13-module framework.

Roadmap

• More markets (Japan, Europe)
• Optional memory layer with manual outcome feedback
• Alternative agent framework integrations (OpenCode, Cursor, generic Python)

Contributing

PRs welcome for: bug fixes in prompts, new markets, additional frameworks, translations.

License

MIT — see LICENSE.

Disclaimer

Research and educational use only. Not investment advice. No guarantee of accuracy or profitability. Consult a licensed professional before making investment decisions.

---

中文说明

这是什么

Claude Code 里一个 /analyze NVDA 命令 → 机构级深度研报，包含：

• 13 模块分析 — 大盘环境、基本面、管理层质量、盈利动量、估值（DCF / 可比 / SOTP / 反向 DCF）、行业定位、催化剂、技术面 + 机构资金、仓位风险、资产负债表、跨资产相关性、泡沫检测、舆情
• 模块十：多智能体多空辩论（3 轮：开场 / 反驳 / 终陈）由组合经理 agent 裁决
• 自我复盘 — 读取同股票过往报告，判断过去预测对错，喂给本次辩论做偏误校正
• 高盛风格 HTML 研报 — 21 区块，高盛配色，A4 打印友好
• 市场覆盖：美股 / 港股 / A 股

vs TradingAgents

	multiagent-stock-research	TradingAgents
分析深度	13 个模块	4 个分析师角色
辩论轮数	3 轮（开场 / 反驳 / 终陈）	可配置，默认 1 轮
市场覆盖	美股 + 港股 + A 股	仅美股
自我反思	有 — 读取自身过往报告	无
输出格式	高盛风格 HTML + 评分卡 + 止损触发条件	交易决策字符串
方法论框架	格雷厄姆 / 巴菲特 / 欧尼尔 / 米纳维尼 / 德鲁肯米勒 / 马克斯	通用 LLM 提示词
安装方式	1 个 shell 脚本，无需 Python 环境	Python 包 + Docker
运行时依赖	无（仅 Claude Code）	LangGraph + 多个 LLM SDK

快速开始

# 1. 克隆
git clone https://github.com/YOUR_USER/multiagent-stock-research
cd multiagent-stock-research

# 2. 安装
./install.sh

# 3. 配置（加到 ~/.zshrc 或 ~/.bashrc）
export TUSHARE_TOKEN="..."      # A 股分析师数据 — https://tushare.pro
export FINNHUB_TOKEN="..."      # 美股内部人/机构数据 — https://finnhub.io
export RESEARCH_OUTPUT_DIR="$HOME/equity-research"   # 可选，默认 ~/equity-research

# 4. 在 Claude Code 中使用
/analyze NVDA --html

命令参数

命令	行为
/analyze TICKER	完整 13 模块 + 辩论，Markdown 输出
/analyze TICKER --html	同上并生成 HTML 研报
/analyze TICKER --silent	静默模式，不输出 Markdown，仅进度行
/analyze TICKER --quick	跳过模块十辩论（快速筛查模式）
/analyze TICKER --lang en	输出英文研报（默认中文）
/analyze T1,T2,T3 --html --silent	批量静默模式
/report TICKER	从最新分析重新生成 HTML

示例报告

真实研报样本：samples/AAPL-sample-report.html，浏览器打开可看完整高盛风格布局，含 Module 10 辩论区（多空 Top 3、核心分歧矩阵、未解风险、含辩论调整的评分卡）。

不克隆也能预览：https://raw.githack.com/chenhab03/multiagent-stock-research/main/samples/AAPL-sample-report.html

英文版 sample（`--lang en` 生成）：samples/AAPL-sample-english.html — 预览: https://raw.githack.com/chenhab03/multiagent-stock-research/main/samples/AAPL-sample-english.html

方法论

详见 docs/methodology.md。

核心框架集成：

• 格雷厄姆（Graham） — 安全边际、Graham Number
• 巴菲特（Buffett） — 护城河 5 维度评分
• 欧尼尔（O'Neil） — CAN SLIM 7 维度评分卡
• 米纳维尼（Minervini） — Stage 分析 + VCP（波动收缩形态）
• 德鲁肯米勒（Druckenmiller） — 宏观 + 集中持仓
• 马克斯（Marks） — 周期定位（债务周期 / 情绪周期 / 资本开支周期）

License

MIT

免责声明

仅供研究和教育用途。不构成投资建议。 过往业绩不保证未来收益，投资前请咨询持牌专业人士。