MCP server for DataSinking — full-text financial reports from the US, China A-shares, Japan, Korea and Taiwan (annual / semi-annual / quarterly) as clean Markdown, ready for LLM reading and RAG.

Turns "what were Toyota's FY2025 results?" into a real answer: the agent finds the filing on EDINET, pulls just the chapter it needs, and cites the source.

npx -y datasinking-mcp

No Python, no uv, no build step — just Node 18+.

Setup

Get a free API key at https://datasink.ing (free tier: 3 requests/second, thousands of documents per month), then add this to your MCP client:

{
  "mcpServers": {
    "datasinking": {
      "command": "npx",
      "args": ["-y", "datasinking-mcp"],
      "env": {
        "DATASINK_API_KEY": "<your-key>"
      }
    }
  }
}

Works with Claude Desktop, Claude Code, Cursor, Devin Desktop, Codex, and anything else that speaks MCP over stdio.

Prefer the remote endpoint? DataSinking also serves MCP over streamable HTTP at https://api.datasink.ing/mcp — no local process, same six tools. Prefer Python? pip install "datasinking[mcp]" gives you the identical server.

Tools

Tool	What it does
`list_exchanges`	Coverage per exchange (`nyse` / `nasdaq` / `cboe` / `sse` / `szse` / `bj` / `jpx` / `ksc` / `koe` / `knx` / `twse` / `tpex`) with report counts
`list_stocks`	Companies on an exchange, with report count each
`list_reports`	A company's reports — metadata only (id, title, period, `source`)
`list_sections`	Every section of one report, with `chars` / `estimated_tokens` / `has_tables`
`get_section`	One chapter, by heading keyword — the token-cheap path for RAG
`get_report`	A whole report (metadata + Markdown body) — expensive, use sparingly

Symbols are FMP-style

AAPL (Apple) · 600519.SS (Kweichow Moutai) · 7203.T (Toyota) · 005930.KS (Samsung Electronics) · 2330.TW (TSMC)

Chapter-level access is the point

A full annual report can be 500k+ characters. list_sections tells the agent what's inside and what it will cost in context, so it pulls one chapter instead of the whole document:

// 1. list_reports   → { "symbol": "7203.T" }              → id = 12345
// 2. list_sections  → { "document_id": 12345 }            → sections: ["事業の概況", ...]
// 3. get_section    → { "document_id": 12345, "section": "事業の概況" }

section is substring-matched against the report's own headings, so pass it in the report's language: 管理层讨论与分析 for A-shares, 事業の概況 for Japanese. If nothing matches, the API returns 404 together with the real headings — retry with one of those.

Data sources

Every response carries a source field naming the official disclosure platform — keep that attribution when you cite a report.

Market	Source	Coverage
US	SEC EDGAR	NYSE / NASDAQ / CBOE
China A-shares	cninfo.com.cn (巨潮资讯网)	5,500+ companies
Korea	DART (전자공시시스템)	KOSPI / KOSDAQ / KONEX
Japan	EDINET (金融庁)	from 2016 — annual, semi-annual and quarterly securities reports (有価証券報告書). Note: 決算短信 are filed on TDnet and are not in EDINET, so they are not here either.
Taiwan	MOPS (公開資訊觀測站)	TWSE / TPEx

Configuration

Env var	Default	Purpose
`DATASINK_API_KEY`	(required)	Your key. Without it every tool call returns a clear error naming this variable.
`DATASINK_API_URL`	`https://api.datasink.ing`	Override the API base URL (self-hosted / testing).

Develop

cd npm
npm install
npm run smoke                                  # handshake + tools/list, no key needed
DATASINK_API_KEY=xxx npm run smoke             # full round-trip against the live API

The smoke test drives this server with a real MCP client over stdio rather than calling the functions directly — tool descriptions and parameter schemas only matter if they actually reach the client, and that has silently broken here before.