DataSinking

DataSinking

[PyPI version] [MCP server]

Full-text financial reports across the US, Europe and Asia, as clean Markdown.

DataSinking serves full-text financial reports — annual, semi-annual and quarterly — from the US, China, Europe, Japan, Korea and Taiwan as clean Markdown, ready for LLM reading and RAG. Query by FMP-style symbol (AAPL, 600519.SS, 7203.T, 005930.KS, 2330.TW) or filter by exchange, report period, or section — pull just the MD&A / risk section instead of the whole filing, saving tokens. Every figure is traceable to its source disclosure and block. Reports are sourced from official disclosure platforms and parsed into structured Markdown with YAML frontmatter, preserved headings, paragraphs and tables.

MCP server

Ship DataSinking to any AI agent (Claude / Cursor / Codex / Devin Desktop / n8n) as an MCP server — 8 tools: list exchanges, list stocks, list reports, fetch a report, list sections, fetch one section (token-friendly for RAG), request a full backfill of a stock (idempotent — returns the current status if already queued), check a backfill's progress.

Hosted — nothing to install

One URL, no package, no Python, no local server. Most clients take this shape:

{  "mcpServers": {    "datasinking": {      "type": "http",      "url": "https://api.datasink.ing/mcp",      "headers": { "Authorization": "Bearer YOUR_KEY" }    }  }}

Claude Code, in one line:

claude mcp add --transport http datasinking https://api.datasink.ing/mcp \  --header "Authorization: Bearer YOUR_KEY"

A client that can't set a header can put the key in the URL instead — https://api.datasink.ing/mcp?apikey=YOUR_KEY — but only do that when you must: a key in a URL ends up in logs and screen shares, and no client can redact it.

Two clients don't fit the shape above: Codex's config is TOML, and Devin Desktop's remote field is serverUrl, not url. Per-client files: docs/mcp/.

Local — run it yourself

If you'd rather keep everything on your own machine, there are two identical builds — pick whichever runtime you already have:

Node 18+ (no Python needed):

{  "mcpServers": {    "datasinking": {      "command": "npx",      "args": ["-y", "datasinking-mcp"],      "env": { "DATASINK_API_KEY": "YOUR_KEY" }    }  }}

Python 3.8+:

pip install "datasinking[mcp]"datasinking-mcp          # requires DATASINK_API_KEY (free at https://datasink.ing)

Then use command: datasinking-mcp in your client.

Both run the same eight tools with the same schemas — npm/ and datasinking/mcp_server.py are kept in lockstep by check_mcp_parity.py.

Full per-client setup: mcp-server.md (overview) · docs/mcp/ (one guide per client).

The hosted endpoint is POST-only and stateless — GET /mcp returns 405, and no session id is issued. Most clients read MCP config at launch, so restart one after changing its config.

[DataSinking MCP in Claude]

What this repo is

Examples, research and tutorials showing how to work with financial report data, including reproducing the presentation styles found in financial-report research papers.

datasinking/├── examples/     # Example scripts: pull data from the API and analyze it├── research/     # Research notes / blog posts (reproducing paper-style presentation)├── datasinking/  # Python client + MCP server — pip install "datasinking[mcp]"├── npm/          # The same MCP server on npm — npx -y datasinking-mcp (Node 18+)├── r-datasinking/  # R client — remotes::install_github("heubme2020/datasinking", subdir="r-datasinking")├── docs/mcp/     # Per-client MCP setup — Claude Code, Desktop, Cursor, Codex, WorkBuddy, Devin…├── mcp-server.md # MCP server overview — endpoint, per-client table, tools, troubleshooting├── llm-examples.md  # Ask an LLM — no code needed (10 end-to-end examples)├── api-examples.md  # 8 examples × 3 interfaces (curl / Python / LLM)└── README.md

Per-client MCP guides — one file each, with the exact config path, both scopes, verification and the errors that client actually produces:

Quick start

1. Get an API key at datasink.ing
2. One line (FMP-style ?apikey=):

curl "https://api.datasink.ing/documents?symbol=600519.SS&with_content=1&apikey=YOUR_KEY"

Or in Python:

pip install datasinking

from datasinking import DataSinking
ds = DataSinking("YOUR_KEY")for r in ds.get_stock_reports("600519.SS", limit=3):    print(r["report_period"], r["title"], len(r["content"]), "chars")

The full reference (curl / Python / LLM): api-examples.md.

Ask an LLM (no code)

Don't want to write code? Point any LLM at datasink.ing, give it your API key, and ask in plain language. See llm-examples.md for ten end-to-end examples — explore coverage, list a company's reports, and extract a figure with correct units.

RAG frameworks

Official integrations load reports straight into LlamaIndex, LangChain, or Haystack — one import, one call:

from datasinking.llama_index import DataSinkingReaderdocs = DataSinkingReader(api_key="YOUR_KEY").load_data("AAPL", doc_type="annual")
from datasinking.langchain import DataSinkingLoaderdocs = DataSinkingLoader(api_key="YOUR_KEY", symbol="AAPL").load()
from datasinking.haystack import DataSinkingFetcherdocs = DataSinkingFetcher(api_key="YOUR_KEY", symbol="AAPL").run()["documents"]

Pass split_by_section=True to load one Document per section (token-friendly for RAG). Each Document's metadata keeps the official disclosure source. Install the client with pip install -U datasinking; each framework is an optional dependency you install separately.

Examples (examples/)

Every example pulls from the live API and runs as-is.

03_download_exchange.py fetches every report on an exchange (e.g. all of Shenzhen — 150k+ documents). Quotas count documents, not requests, over a rolling 7-day window: a free key gets 3 req/s and 8,191 documents per 7 days, inside a pool of 524,287 per 7 days shared by all free users and website visitors. A whole exchange will therefore take well over a week on a free key — a paid (yearly) key (31 req/s, 524,287 documents per 7 days) is strongly recommended.

Research (research/)

research/ hosts research notes and blog posts, each based on DataSinking data with the source cited. You can reproduce charts and presentations found in financial-report research papers, e.g.:

• Long-term revenue / profit trends
• Industry comparison and distribution
• Time series of financial metrics

Start from research/TEMPLATE.md.

Data overview

Data source

Reports are sourced from the official regulatory disclosure platform of each market and converted in-house to clean Markdown:

Every document also carries a source field in the API response, so the attribution travels with the data. Please keep it when you redistribute.

How to cite

If you use DataSinking in a paper, add a data-availability note (and, optionally, an acknowledgment).

Data availability:

Full-text financial reports (annual, semi-annual, and quarterly filings across the US, China, Europe, Japan, Korea, and Taiwan) were obtained from DataSinking (https://datasink.ing) in Markdown format.

BibTeX:

@misc{datasinking2026,  author       = {{DataSinking}},  title        = {{DataSinking: Full-text financial reports as Markdown}},  year         = {2026},  howpublished = {https://datasink.ing},  note         = {Accessed: 2026-10-27}}

DataSinking provides report text, not investment advice.

Affiliate

Earn $7 every purchase for every customer you refer — recurring, PayPal, $31 minimum withdrawal: datasink.ing/affiliate

License

MIT