// the find
simonlin1212/a-stock-data
A股全栈数据工具包：行情K线·当日逐笔·研报·信号·资金面·新闻·财务·公告·打板·ETF期权·舆情·宏观利率·期货大宗(含大商所日K)·事件驱动·可转债 | 15层·87端点·34数据源·除iwencai外免Key | A-share data for AI agents: K-lines, ticks, reports, fund flow, news, financials, filings, options, macro, futures, events, convertibles | 15 layers·87 endpoints·34 sources
A single self-contained SKILL.md file (plus a thin Python layer) that scrapes 87 endpoints across 34 Chinese financial data sources — quotes, K-lines, financials, options, futures, convertible bonds, macro rates, filings — built specifically to be dropped into Claude Code or similar agent contexts. It's for developers building AI agents or quant tooling around the A-share market who don't want to hand-roll auth and parsing for a dozen different Chinese finance sites.
Distribution as one markdown file with embedded Python is a genuinely good fit for its stated use case (agent context injection) — copy one file, it works, no package versioning to manage. The changelog and FAQ show real production debugging: documented unit bugs (ShenZhen 688/689 board volume being shares vs lots), a corrected PB field index, and a dead upstream (mootdx TCP quotes returning empty since Sept 2026) with a concrete replacement path rather than silently breaking. 173 offline tests extract and run the code blocks straight from the SKILL.md, so the docs and the code can't drift apart undetected. Source prioritization is deliberate, not accidental — Tencent/exchange-official first because they don't rate-limit, Eastmoney last because it does, which shows the author actually hit these walls.
**Weaknesses:**
Every endpoint is an unauthenticated scrape of undocumented internal APIs (Tencent, Eastmoney push2/datacenter, Sina), which means there's no contract — any of the 34 sources can change response shape or get blocked with zero warning, and the FAQ already lists several that intermittently do. The project isn't a Python package; there's no `pip install a-stock-data`, no types, no API surface beyond "paste this markdown into your LLM's context," so it's unusable as a normal library dependency in a non-agent codebase. Correctness depends on mainland China network egress for several sources (mootdx needs a domestic IP, Eastmoney rate-limits foreign/cloud IPs specifically) — the FAQ explicitly tells overseas users to proxy or switch tools, which is a real tax on anyone not running from a CN IP. No backtesting engine and explicitly no point-in-time historical index constituents, so despite the huge endpoint count this is a data pipe, not a research platform — you still have to build everything downstream yourself.