Disclosure Compass

A focused Model Context Protocol server for public company disclosures from OpenDART (전자공시시스템 DART). Disclosure Compass (공시나침반) exposes ten read-only gateway tools over Streamable HTTP. Behind that compact public surface are 16 in-process specialist MCP servers with 82 OpenDART tools, distilled from the original implementation without its LLM, vector-search, or A2A runtime dependencies.

This project is independent open-source software and is not an official OpenDART product.

Tools

Tool	Purpose
get_company_profile	Basic company and listing information
search_disclosures	Disclosure filings in a bounded date range
get_financial_statement	A bounded set of statement accounts
get_dividend_information	Current and prior-period dividend rows
get_major_shareholders	Major shareholder positions
get_employee_statistics	Employee counts, tenure, and pay statistics
classify_disclosure_request	Rank a Korean request across 16 disclosure domains and show their specialist tools
list_disclosure_servers	Inspect all 16 specialist servers and 82 tool names/endpoints
call_disclosure_server_tool	Execute an exact named tool on an exact specialist server
route_and_call_disclosure	Classify, select a specialist tool, and execute it in one call

All tools are declared read-only, non-destructive, idempotent, and open-world. Every specialist result is capped at 50 rows. Binary filing and XBRL tools return bounded ZIP metadata rather than raw file contents. Upstream calls use strict connect/read timeouts and reject JSON responses larger than 2 MB or binary responses larger than 8 MB.

Disclosure request classification

classify_disclosure_request maps a non-empty Korean natural-language request of up to 500 characters to the most relevant OpenDART disclosure domains. top_k defaults to 3 and may be set from 1 through 16. The response contains the original query, total_categories, and score-ranked routes. Each route contains:

• category: the stable domain ID
• label_ko: the Korean domain label
• score: the normalized relevance score
• matched_terms: terms from the request that contributed to the route
• supported_tools: data tools currently exposed by this server for the domain
• specialist_tools: original tool names available on that domain's MCP server

This classifier performs deterministic term matching and does not itself call OpenDART. Use route_and_call_disclosure to classify and execute in one request, or use call_disclosure_server_tool for deterministic server/tool selection. The 16 specialist servers are real FastMCP instances called through in-memory FastMCP clients; they are not 16 separately deployed network services.

The classifier recognizes exactly these 16 domains:

Category	Korean label
disclosure_search	공시검색/기업개황
shareholder_stock	주주/주식 정보
executive_compensation	임원/보수 정보
debt_securities	채무증권 정보
audit_fund	감사/자금 정보
financial_statement	재무정보
equity_disclosure	지분공시
capital_change	증자감자
treasury_stock	자기주식
convertible_securities	전환증권
merger_division	합병분할
business_transfer	영업/자산 양수도
overseas_listing	해외상장
equity_investment	지분거래
corporate_issues	기업이슈
securities_registration	증권발행/등록정보

Example request:

{
  "query": "연결 재무제표 매출액",
  "top_k": 1
}

Example response:

{
  "query": "연결 재무제표 매출액",
  "routes": [
    {
      "category": "financial_statement",
      "label_ko": "재무정보",
      "score": 1.0,
      "matched_terms": ["연결 재무제표", "재무제표", "매출액"],
      "supported_tools": ["get_financial_statement"],
      "specialist_tools": [
        "dart_fnlttSinglAcnt",
        "dart_fnlttMultiAcnt",
        "dart_fnlttXbrl",
        "dart_fnlttSinglAcntAll",
        "dart_xbrlTaxonomy",
        "dart_fnlttSinglIndx",
        "dart_fnlttCmpnyIndx"
      ]
    }
  ],
  "total_categories": 16
}

To execute a routed request directly, pass a company identifier and any period-specific fields to route_and_call_disclosure:

{
  "query": "삼성전자 전환사채 발행 결정을 찾아줘",
  "corp_code": "00126380",
  "begin_date": "20250101",
  "end_date": "20251231"
}

This selects the convertible_securities server and its dart_cvbdIsDecsn tool. Callers that already know the exact target can instead use call_disclosure_server_tool with server_id, tool_name, and an arguments object. Accepted fields depend on the endpoint and include corp_code or corp_name, business_year, report_code, begin_date, end_date, corp_codes, receipt_number, fs_division, statement_type, index_code, and max_items.

Requirements

• Python 3.11-3.13
• An OpenDART API key from https://opendart.fss.or.kr/

Run locally

python -m venv .venv
. .venv/bin/activate
pip install -e '.[dev]'
export DART_API_KEY='your-runtime-secret'
opendart-mcp

The MCP endpoint is http://localhost:8000/mcp and the health endpoint is http://localhost:8000/health. Set PORT to override port 8000.

The corp_code arguments are OpenDART eight-digit company identifiers, not six-digit stock codes. For example, OpenDART's public guide uses 00126380 for Samsung Electronics.

Docker / PlayMCP deployment

The image is compatible with Linux AMD64 and runs as a non-root user.

docker build --platform linux/amd64 -t opendart-mcp .
docker run --rm -p 8000:8000 \
  -e DART_API_KEY='your-runtime-secret' \
  opendart-mcp

Configure the deployment secret as DART_API_KEY; never bake it into the image or commit it. Register /mcp as the Streamable HTTP endpoint. The server uses stateless HTTP so it does not require session affinity.

Development checks

ruff check .
pytest -q
python -m compileall -q src tests

Tests use local fixtures and never call OpenDART or require an API key.

Data and operational notes

• Results are public disclosure data, but users should confirm material facts in the original filing before making decisions.
• OpenDART may update data between otherwise identical calls.
• The server does not store query inputs or results.
• Upstream rate limits and maintenance windows still apply.

License

MIT. See LICENSE.