Korea Data Suite MCP Server

Korea Data Suite

English | 한국어

Clean, developer-friendly REST APIs for Korean public data. Korean government open data is powerful but hard to consume — Korean-only docs, XML responses, legacy auth. This suite normalizes it into simple JSON APIs.

▶ Try it on RapidAPI — free tier, no setup. Hosted and auto-updated; same code as this repo. For AI agents, it's on the MCP Registry — uvx korea-data-mcp.

APIs

Get started

Hosted (recommended) — a maintained instance with a free tier and no setup:

→ Subscribe on RapidAPI, grab your key, and call any endpoint. RapidAPI injects the key for you — copy a ready-made snippet from its Code Snippets panel.

Both run the exact same code (this repo). Pick RapidAPI if you'd rather not operate data pipelines; self-host if you want full control.

Self-host

uv sync
uv run uvicorn app.main:app --port 8642
curl "http://127.0.0.1:8642/v1/health"

Holidays & Business Days API

# All holidays in a year (or a month)
curl "http://127.0.0.1:8642/v1/holidays?year=2026" -H "X-API-Key: <key>"

# Is a given date a holiday / business day?
curl "http://127.0.0.1:8642/v1/holidays/check?date=2026-03-02" -H "X-API-Key: <key>"

# Add N business days (skips weekends & holidays)
curl "http://127.0.0.1:8642/v1/business-days/add?date=2026-12-31&days=1" -H "X-API-Key: <key>"

# Count business days in a range (inclusive)
curl "http://127.0.0.1:8642/v1/business-days/count?start=2026-09-21&end=2026-09-27" -H "X-API-Key: <key>"

Covers official public holidays, substitute holidays (대체공휴일), temporary holidays (임시공휴일), and election days — the cases most global holiday APIs get wrong for Korea.

Real Estate Transactions API

Normalized MOLIT (Ministry of Land) real transaction prices — apartment, officetel, and land; sale, jeonse, and monthly-rent — as clean English JSON with cursor pagination.

# Real transaction prices (apartment sales in Gangnam-gu)
curl "http://127.0.0.1:8642/v1/realestate/transactions?region=11680&property_type=apartment&trade_type=sale" -H "X-API-Key: <key>"

# Filter by date range + paginate with the returned cursor
curl "http://127.0.0.1:8642/v1/realestate/transactions?region=11680&date_from=2026-01-01&limit=50&cursor=<next_cursor>" -H "X-API-Key: <key>"

# Region codes (LAWD 5-digit)
curl "http://127.0.0.1:8642/v1/realestate/regions" -H "X-API-Key: <key>"

Daily sync ingests the current + previous month; use the backfill CLI for history:

uv run python scripts/backfill.py --from 2025-01 --to 2025-12 --regions 11680,11650

Configuration

Environment variables (prefix KDS_, .env supported):

Data sources & attribution

• Holiday data: KASI Special Day Information (한국천문연구원 특일정보), via Korea Public Data Portal (data.go.kr) — KOGL Type 1. Ships with bundled seed data (2025–2027); refreshed weekly when a service key is configured.
• Real transaction data: MOLIT 실거래가 공개시스템 (국토교통부), via Korea Public Data Portal (data.go.kr) — KOGL Type 1.

Run as a daemon (macOS)

# Install & start (auto-restart on crash, start at login)
./scripts/install-daemon.sh

# With Cloudflare Tunnel (after one-time `cloudflared tunnel login/create`)
./scripts/install-daemon.sh --with-tunnel

# Logs
tail -f ~/Library/Logs/kds/api.out.log

# Uninstall
launchctl bootout "gui/$(id -u)" ~/Library/LaunchAgents/com.choiyounggi.kds-api.plist
rm ~/Library/LaunchAgents/com.choiyounggi.kds-api.plist

To keep the machine awake for serving, disable system sleep (sudo pmset -a sleep 0) or use a dedicated always-on machine. See deploy/cloudflared.example.yml for exposing the API via Cloudflare Tunnel without opening ports.

Handling concurrent traffic

The read path and the write path are separated so traffic scales independently:

• SQLite in WAL mode (set once at init) + busy_timeout — readers never block the daily writer and vice-versa, and multiple read workers can run concurrently.
• API process is read-only, multi-worker. scripts/run.sh runs uvicorn with --workers ${KDS_WORKERS:-2} and KDS_ENABLE_SCHEDULER=false. Each worker is a separate process (separate GIL); WAL lets them all read at once. Raise KDS_WORKERS to scale reads with cores.
• The daily ingest runs as its own process (com.choiyounggi.kds-sync, 04:00) via scripts/sync.py — never inside the API server, so a multi-thousand-row batch never competes with request handling for the GIL.
• Edge caching (optional): responses carry Cache-Control: no-store for security. The real-estate data is public and changes at most daily — if origin load grows, serve it with a short Cache-Control: public, max-age=... and let the CDN absorb reads.

Security checklist before exposing externally

The app is hardened at the code layer (API-key auth fail-closed, parameterized SQL, strict input validation, security headers on every response including 5xx, docs/schema off by default, sanitized errors). The following are edge/deploy responsibilities that must be in place before opening the tunnel:

• Never set KDS_DEV_MODE=true in production — it disables all auth. The app logs a warning at startup if it is on.
• Cloudflare rate limiting + WAF on the tunnel hostname — the app has no app-layer rate limit by design (edge responsibility).
• HSTS + TLS are terminated at the Cloudflare edge; confirm HSTS is enabled there (the origin serves plain HTTP on 127.0.0.1 only).
• Keep KDS_ENABLE_DOCS unset (or false) in production; set true only to serve /docs /openapi.json at the origin.

SEO marketing site (programmatic)

A static, SEO-optimized marketing site is generated from the live DB by scripts/gen_site.py. For every region that has real transaction data it emits a Korean landing page (the query users actually type — "강남구 아파트 실거래가 API" — backed by real MOLIT stats, a working curl example, and a signup CTA), plus a holidays pillar page, a home page, sitemap.xml, and robots.txt.

Quality gate (important): a region is only published if it has at least MIN_SALE_ROWS (30) apartment-sale rows. Regions without enough data are skipped — this deliberately avoids thin/doorway pages, which search engines penalize.

# generate into site/dist (reads data/kds.db)
uv run python scripts/gen_site.py --out site/dist

Config is env-driven so the same generator works for any domain (put these in deploy/site.env, gitignored — copy deploy/site.env.example):

Serving — the API app serves it

The FastAPI app serves site/dist at all non-API paths (app.mount("/")), while /v1/* stays the JSON API. The two get different response headers: the API keeps its locked-down default-src 'none' CSP + no-store; the site gets an HTML-renderable CSP (script-src 'none', inline styles allowed) + public cache. Files are read from disk per request, so regenerating the site goes live with no app restart — only a code change needs a restart.

The site is served on the same host as the API (api.korea-data.cloud) — the API lives under /v1, the site everywhere else — so no new tunnel hostname or DNS is needed. One-time on the serving host:

cp deploy/site.env.example deploy/site.env    # KDS_SITE_URL == KDS_API_ORIGIN == https://api.korea-data.cloud
uv run python scripts/gen_site.py --out site/dist   # generate once
# restart the API app so this integration (new code) takes effect — the site is
# then live at https://api.korea-data.cloud/ , /holidays/ , /realestate/... .

Submit https://api.korea-data.cloud/sitemap.xml once in Google Search Console.

Want the site on a bare korea-data.cloud / www later? Add an ingress rule pointing that hostname at the same http://127.0.0.1:8642, route its DNS, and switch KDS_SITE_URL to it. Not required — the api host works for SEO today.

First run needs history: the daily sync only ingests the current month. To give pages real depth, backfill once — uv run python scripts/backfill.py --from 2025-07 --to 2026-06 --regions <codes> --datasets apt_trade,apt_rent.

Automate (macOS daemon)

deploy/com.choiyounggi.kds-site.plist regenerates the site daily at 04:30 (right after the 04:00 sync) via scripts/publish_site.sh. Because the app serves from disk, the refreshed pages are live immediately — no restart, no external deploy:

cp deploy/com.choiyounggi.kds-site.plist ~/Library/LaunchAgents/
launchctl bootstrap "gui/$(id -u)" ~/Library/LaunchAgents/com.choiyounggi.kds-site.plist
tail -f ~/Library/Logs/kds/site.out.log

Uptime note: since the site is served by the local app (not a CDN), its SEO availability tracks the machine — keep it awake for serving (pmset, as the API already requires). If always-on hosting is wanted later, the same site/dist can be pushed to Cloudflare Pages instead.

MCP server (AI-agent access)

The API is also packaged as a standalone Model Context Protocol server — packages/korea-data-mcp/ — so AI agents (Claude Desktop/Code, Cursor, …) can discover and call the endpoints directly. It is its own minimal package (deps: mcp, httpx) and is what gets published to PyPI / listed in the MCP Registry.

Quick add (before PyPI, straight from this repo — users bring their own key):

claude mcp add korea-data-suite --env KDS_API_KEY=<key> \
  -- uvx --from "git+https://github.com/choiyounggi/korea-data-suite#subdirectory=packages/korea-data-mcp" korea-data-mcp

See packages/korea-data-mcp/README.md for tools, client config, and uvx korea-data-mcp (once on PyPI).

License

MIT © choiyounggi