Claude Code-driven Stock Research Agent: A Quantitative Analysis Workflow Without Paid APIs

This article introduces an innovative agentic workflow for stock research: using Claude Code as the intelligent layer and GitHub Actions to automate stock screening, evidence collection, and analytical decision-making. It builds a high-confidence tech stock selection system without LangGraph or paid APIs. Key advantages include zero API costs, no complex orchestration frameworks, a transparent and auditable decision process, providing professional-grade research tools for individual investors and small teams.

Traditional quantitative investment solutions rely on expensive data APIs and complex orchestration frameworks. This project adopts "simplicity first" as its core design philosophy: no dependence on frameworks like LangGraph, no use of paid APIs, and full use of Claude Code's agent capabilities to complete analysis tasks. It focuses on Nasdaq and Indian NSE tech stocks, screening high-beta stocks with a market cap over $200 million; runs every Thursday evening, producing 0-5 high-confidence recommendations. An "empty inbox" is considered normal (strictly controlling investment quality).

First Stage: Deterministic Screener

Triggered regularly via GitHub Actions (every Thursday evening at 9 PM Indian time), it executes dependency installation, unit tests, and stock screening scripts, outputting CSV files of candidate/rejected stocks (with reasons). It covers about 1500 US and NSE stocks, taking 30-45 minutes.

Second Stage: Evidence Collection Toolset

Provides 8 functional functions (e.g., SEC filings, company news, earnings call analysis, etc.). Return values include a list of sources and source classification (Tier A: official documents; Tier B: mainstream media; Tier C: low-quality sources are directly discarded).

Third Stage: Three-Question Analyst (ReAct Loop)

Answers three questions for each of the top 25 candidate stocks: 1. Is the market pricing correct? 2. Is the mismatch temporary or structural? 3. What is the biggest risk? Each question can call tools up to 12 times to ensure evidence support.

Evidence Collection Tools

Covers 8 dimensions: SEC documents, company news, earnings meetings, analyst ratings, insider transactions, peer reactions, governance risks, and macro background. Source classification ensures credibility.

Rule-Driven Critic Mechanism

Makes final decisions based on comprehensive analysis results: PICK (high-confidence recommendation), WATCH (needs observation), SKIP (does not meet standards). Decision trails are written to markdown files, allowing traceability of the reasons for selecting/rejecting each stock.

1. Zero API Cost: Data comes from free tiers of yfinance, SEC EDGAR, and Finnhub; the intelligent layer uses Claude Code (developer subscription) with no additional costs.
2. No Orchestration Framework: Claude Code itself serves as the agentic layer, simplifying system complexity and improving maintainability.
3. GitHub Actions Automation: The first stage runs fully automatically on a schedule and supports manual triggering.
4. Auditable Records: All results are saved as markdown/CSV files, forming historical archives for easy traceback and strategy optimization.

Usage Flow

1. Thursday evening: GitHub Actions automatically executes screening; 2. The user inputs "run weekly analysis" to Claude Code; 3. AI completes in-depth analysis of 25 stocks; 4. Check the output in the runs directory.

Investment Insights

• Efficiency improvement: Reduces research time and focuses on high-value judgments;
• Enhanced discipline: Rule-driven to eliminate emotional decisions;
• Knowledge precipitation: Analytical logic and basis form reusable assets;
• Low-cost threshold: Individuals/small teams can use professional tools.

Limitations

• Data coverage: Only US and Indian tech stocks; needs expansion to other markets/industries;
• Real-time performance: Dependence on free data sources may lead to delays;
• Model dependence: Differences in Claude Code versions affect analysis quality;
• Lack of backtesting: No built-in backtesting framework to verify strategy effectiveness.

Improvement Directions

• Expand market coverage (A-shares, Hong Kong stocks, Europe);
• Integrate backtesting functions;
• Add portfolio management modules;
• Develop a visual interface to lower the threshold.

This project demonstrates the possibility of building practical AI systems with limited resources, proving that complex financial analysis does not require expensive APIs or frameworks. As an open-source example, it provides an excellent starting point and reference for developers and investors exploring AI applications in the investment field, helping to lower the threshold for using professional research tools.