Korean DART MCP

83 OpenDART APIs packed into 15 MCP tools. Disclosure search, financial statements, shareholder structure, XBRL, analyst frames (insider signals, governance risk scoring, Buffett-grade quality checklists), and HWP/PDF attachment-to-markdown conversion — all usable directly from any AI assistant.

MCP server + CLI wrapping OpenDART (Financial Supervisory Service's public disclosure platform — Korea's equivalent of the SEC's EDGAR). Works with Claude Desktop, Cursor, Windsurf, Claude Code, and any MCP-compatible client.

Sister project: korean-law-mcp (Korean statute database, 41 APIs → 15 tools).

한국어 문서 → README.md

---

Why this exists

Korea has ~3,000 listed companies whose filings, financials, ownership data, and XBRL flow through DART (Data Analysis, Retrieval and Transfer System). Two mature Python wrappers — OpenDartReader (438⭐) and dart-fss (364⭐) — already map the 83 raw endpoints for pandas users.

This project targets a different layer:

• For pandas users — OpenDartReader / dart-fss. Analysts wrangle DataFrames directly.
• For LLM-native users — this project. Raw tables get refined into the angles an expert actually uses (Buffett checklists, insider trade cluster signals, governance risk scores, capital-event timelines, markdown full-text) so an AI agent can build a narrative on first pass.

They complement each other. Want DataFrames? Use the Python wrappers. Want agent-ready frames? Use this MCP.

---

v0.9 — What's new

• get_xbrl format="markdown_full" — full presentation/calculation linkbase parsing: every account with hierarchy + calculation-linkbase validation. BS 50+ / IS 15+ / CF 10+ rows vs v0.8's 50-tag whitelist. Handles industry-specific taxonomies (financial holdings DX prefix, insurance) automatically. 6MB XBRL → ~30-60KB markdown.
• search_disclosures auto-split — no corp_code + range >90 days auto-chunks into 90-day windows (works around OpenDART's "3-month limit for market-wide queries"). Cap: 40 chunks (~10 years).
• summary_text field on insider_signal and disclosure_anomaly — one-line Korean summaries for quick context before raw tables.
• Security hardening (v0.9.1) — ZIP slip / ZIP bomb guards via shared helper, HTTPS viewer scraping, chunk cap, presentation-recursion depth guard, XBRL parse-warning exposure.

Full history → CHANGELOG.md

---

💡 For retail investors — 5 things you can actually do

Practical use cases for Korean-equity retail investors whose broker app isn't enough. Just ask Claude in plain language.

1. "Is management buying with their own money on my holdings?"

"Samsung Electronics — insider buy/sell activity over the last 12 months"

→ insider_signal aggregates reports into cluster signals. Live result: 2,429 buys vs 43 sells → strong_buy_cluster. A 24:1 buying dominance by executives. Impossible to spot manually from DART's filing-by-filing view.

2. "Any accounting skeletons in this company?"

"Kakao — 3-year accounting/governance risk score"

→ disclosure_anomaly returns a 0-100 score + verdict (clean/watch/warning/red_flag). Live: Kakao 40/100 warning — amendment ratio 32.8% exceeds the 20% threshold. Automated flag detection no retail investor does by hand.

3. "I don't want to read a 300-page annual report"

"Summarize the 'Risk Factors' and 'Business Overview' sections from Samsung's 2023 annual report"

→ get_attachments(mode="extract") converts 2.2MB PDFs into 920k-char markdown in 3.7 seconds. Claude reads and summarizes section-by-section. You get primary-source reading without paying for analyst reports.

4. "Who filed what today?"

"All listed companies that filed treasury stock purchase decisions in the last 30 days"
"Convertible/exchangeable bond issuance filings, last 7 days"
"M&A / spin-off decisions in the last 30 days"

→ search_disclosures(preset=...) covers 22 presets. Treasury buys = bullish signal / CB/BW issuance = dilution warning. Live: 59 treasury-buy filings in 30 days. Batch views brokerage apps miss entirely.

5. "Which of these is more solid, A or B?"

"Compare Samsung, SK Hynix, LG Electronics on 5-year ROE / debt / growth"

→ buffett_quality_snapshot(corps=[...]) auto-ranks across 5 metrics. Live (see Scenario 1 above): SK Hynix passes 3/4 checklist; Samsung dominates debt stability. Pick stocks by numbers instead of vibes.

---

Who this fits

• Intermediate retail investors holding 5-20 stocks, reviewing disclosures quarterly/semi-annually
• People who want primary-source disclosures instead of journalist-filtered news
• Investors frustrated by the limits of Naver Finance / HTS app data
• DIY analysts who'd rather not pay for brokerage research

Who this is overkill for

• Day traders on charts — no chart / real-time price here
• KOSPI ETF buy-and-holders — no need for single-stock analysis
• Quants running DataFrames in Excel/Python — use OpenDartReader or dart-fss instead (pandas-native)

Honest barriers to entry

• Claude Desktop / Cursor install + OpenDART key provisioning (~10 min, free, 20k requests/day)
• Terms like ROE / CAGR / debt-to-equity are assumed
• Claude Pro subscription $20/month (large-PDF summarization benefits from a plan with a bigger context window)
• This is a research assistant. Investment decisions are yours.

---

Real scenarios — live API results

All results below are actual values from live DART API calls. Reproducible via scripts/showcase-v0_9_1.mjs (12/12 PASS).

1. Buffett-style 5-year quality comparison + auto-ranking

Prompt: "Compare Samsung, SK Hynix, LG Electronics on 5-year quality metrics"

→ buffett_quality_snapshot(corps=["삼성전자","SK하이닉스","LG전자"], years=5)

Company	Avg ROE	Latest D/E	Revenue CAGR	Net Income CAGR	Checklist
Samsung	10.39%	29.94%	4.51%	3.17%	1/4
SK Hynix	12.86%	45.95%	22.6%	45.37%	3/4
LG Electronics	5.37%	140.33%	4.81%	-3.63%	0/4

Auto-generated rankings (5 metrics): ROE → SK Hynix > Samsung > LG · Debt stability → Samsung > SK Hynix > LG · Net Income CAGR → SK Hynix > Samsung > LG · ROE stability (stddev ↓) → LG > Samsung > SK Hynix.

2. Is management buying with their own money? (insider_signal)

Prompt: "Samsung insider buy/sell cluster analysis, last 12 months"

→ insider_signal(corp="삼성전자", start="2025-04-18", end="2026-04-18")

Samsung Electronics: 2,473 reports (buy 2,429 / sell 43).
1,047 distinct buyers vs 40 sellers. Net +2,302,375 shares.
→ strong_buy_cluster signal.
Peak cluster: 2026Q1 (985 buyers / 18 sellers).

Buffett's "is management buying with their own money?" quantified in one call. 24:1 buying dominance recently.

3. Accounting/governance risk score (disclosure_anomaly)

Prompt: "Kakao last 3 years — accounting risk"

→ disclosure_anomaly(corp="카카오")

Kakao (2023-04 ~ 2026-04): ⚠️ WARNING, score 40/100
- Amendment filings 167/509 (32.8%)  ← over 20% threshold, +30 pts
- Capital stress filings: 5           ← +10 pts
- verdict: warning

Four axes (amendments / auditor churn / non-clean opinion / capital stress) → 0-100 with per-flag evidence. LLM just writes the story.

4. XBRL all-accounts + calculation validation (v0.9)

Prompt: "Samsung 2023 annual — full financial statements"

→ get_xbrl(rcept_no="20240312000736", format="markdown_full")

Periods: current 2023-12-31 / prior 2022-12-31 / prior-prior 2021-12-31
Account rows: BS 52 · IS 18 · CF 12 (vs 17/13/7 in whitelist mode — 3× more)
Markdown size: 8,905 chars (6MB XBRL → 99.85% reduction)
Calc validation: ✅ all balanced (0 violations)
Taxonomy: 10 presentation roles · 8 calculation roles
Elapsed: 615ms

Calc validation is the killer — summation-item relations in the calculation linkbase catch reporting errors on the spot.

5. Industry-specific taxonomy (financial holdings)

Prompt: "Shinhan Financial Group latest annual — full financials"

→ search_disclosures to resolve rcept_no → get_xbrl(format="markdown_full")

Shinhan 2025 annual (rcept_no=20260318000826)
BS 44 rows · IS 49 rows
3 calc-validation violations (financial-industry-specific items)
→ DX-prefix (financial holding) taxonomy handled without code changes

dart-fss downloads XBRL zips but auto-handling of the financial DX prefix isn't documented there.

6. All treasury-stock purchase decisions, last 30 days

Prompt: "All listed firms that filed treasury stock purchase decisions in the last 30 days"

→ search_disclosures(preset="treasury_buy", days=30, limit=500)

Matched: 59 filings / 8 pages in parallel (17.5s)

Latest 5:
  2026-04-17 Tiplax — treasury stock trust contract termination
  2026-04-17 M2N — treasury stock purchase decision
  2026-04-17 PS Electronics — treasury stock trust termination
  2026-04-15 Asia — treasury stock trust termination
  2026-04-15 Asia Cement — treasury stock trust termination

22 presets auto-assemble the pblntf_ty + report_nm regex — the LLM doesn't have to memorize DART codes.

7. Market-wide 180-day query (auto-split, v0.9)

Prompt: "All annual reports filed in the last 6 months, market-wide"

→ search_disclosures(preset="annual_report", days=180)

Auto-split: 3 chunks (bypassing DART's "3-month limit for market-wide queries")
Fetched 6,000 → matched 2,625 annual reports (10.3s)

8. Capital-event timeline (Kakao, 3 years)

Prompt: "All Kakao capital events over the last 3 years"

→ get_corporate_event(corp="카카오", mode="timeline", start="2023-04-18", end="2026-04-18")

Event type	Count
Treasury stock disposal	14
Capital reduction	3
Merger	2
CB issuance	1
EB issuance	1
Total	21

36 event enums fetched in parallel → merged by date.

9. Unified ownership filings (5%-rule + executive stake)

Prompt: "Samsung 3-year ownership changes"

→ get_major_holdings(corp="삼성전자")

majorstock (5%-rule): 41 filings — latest: Samsung C&T 19.70% (2026-04-17)
elestock (executive/major holder): 200 of 2,615 returned (latest first)

One call, two endpoints merged — Python wrappers need two calls + pandas merge.

10. Single-company Buffett checklist with evidence

Prompt: "Samsung 6-year Buffett checklist"

→ buffett_quality_snapshot(corps=["삼성전자"], years=6)

Samsung 2020-2025:
- ROE avg 10.26% (min 4.26 / max 15.69 / stddev 3.58)
- D/E latest 29.94% / avg 31.11%
- Revenue CAGR 7.09% · Net income CAGR 11.35%

Checklist 3/4:
  ❌ consistent_high_roe (all years ROE ≥ 15%)
  ✅ low_debt            (latest D/E ≤ 100%)
  ✅ growing_revenue     (revenue CAGR ≥ 5%)
  ✅ growing_earnings    (net income CAGR ≥ 5%)

11. Filing full text as markdown

Prompt: "Samsung latest treasury-stock decision — full text"

→ search_disclosures(preset="treasury_buy") → download_document(format="markdown")

Original: 2026-03-18 Treasury stock purchase decision
XML 32,618 chars → markdown 2,272 chars (93% reduction)
Headings and tables preserved.

12. Annual-report attachment → markdown

Prompt: "Samsung 2023 annual report PDF body"

→ get_attachments(rcept_no="20240312000736", mode="extract", index=0)

• 2.2MB PDF → 921,998 chars of markdown in 3.7s. LLM can search "Risk Factors" directly.
• kordoc engine (HWP/HWPX/PDF/DOCX/XLSX) — neither OpenDartReader, dart-fss, nor any of the 6 existing DART MCP servers offers this.

---

How this compares to existing DART tooling

Ecosystem survey (as of 2026-04-18):

Feature	OpenDartReader (438⭐, Python)	dart-fss (364⭐, Python)	hypn4/opendart-fss-mcp (85 tools)	RealYoungk/opendart-mcp (83 tools)	korean-dart-mcp (15 tools)
MCP-native	❌	❌	✅	✅	✅
Node.js/TypeScript (npm)	❌	❌	❌	❌	✅ (only one)
1:1 endpoint coverage	most	filings+financials	all 85	all 83	compressed 83→15 via enums
Company name auto-resolve	partial	partial	✅ (typo/consonant)	✅	✅ (SQLite FTS preload)
XBRL presentation/calculation linkbase	❌	ZIP only	ZIP+taxonomy	ZIP only	✅ auto markdown + calc validation
HWP/PDF attachment → markdown	❌	❌	❌	❌	✅ (only one, kordoc)
insider_signal cluster	❌	❌	❌	❌	✅ (only one)
disclosure_anomaly 0-100 score	❌	❌	❌	❌	✅ (only one)
buffett_quality_snapshot checklist	❌	❌	❌	❌	✅ (only one)
90-day auto-split · parallel paging	❌	❌	❌	❌	✅
ZIP slip/bomb hardening	n/a	n/a	❌	❌	✅

Positioning

Python wrappers (OpenDartReader, dart-fss) are for quants/backtesters using DataFrames in Jupyter. The 6 existing Python DART MCPs hand DART's raw JSON to LLMs verbatim. This project is the only LLM-native Node.js MCP in Korea's DART ecosystem — 83 APIs compressed to 15 enum-based tools, with XBRL full parsing, HWP/PDF-to-markdown, and insider/anomaly/Buffett analyst frames built in.

Honest limitations

• 1:1 endpoint coverage is wider on hypn4 (85 tools) and RealYoungk (83 tools) — if you need a rare endpoint directly, those are better choices. This project c