KoreaAPI

The verifiable data layer for Korean culture & commerce, callable by any AI agent. The MCP gateway to Korea — verifiable.

KoreaAPI exposes Korean culture, entertainment, and commerce data to AI agents via Anthropic's Model Context Protocol (MCP). Every response carries machine-readable provenance and a Skill Score so an agent can decide whether to trust and cite it.

Status: Phase 1 (cold-start). The locked spec is in SCOPE.md; what is built / decided and why is in ROADMAP.md. Live, verified, public data (Schema.org JSON-LD + /llms.txt): https://kwangdol-star.github.io/koreaapi/ Repository: kwangdol-star/koreaapi — a standalone repo, split out from its incubation home with full git history preserved.

What's live now (verified, on the public page + via MCP)

• ~120 verified entities across 3 verticals — K-pop artists, K-dramas, K-films — each cross-verified, identity-/hallucination-guarded, and stamped with provenance + a Skill Score.
• A knowledge graph, not a flat list — four linked axes: entities · people (cast / director / member hubs) · verticals (artists / dramas / films) · labels (소속사 / network — "who's under HYBE / on Netflix?"). Every page cross-links to its neighbours.
• Cross-verification — Wikidata + Wikipedia must agree on the canonical bilingual name before a fact clears the single-source cap (high Skill Score = independent concurrence); a strict Korean-name guard rejects same-English-name impostors. Every failure fails safe to a miss.
• Callable by agents (MCP) — 11 tools: trust (get_verified, get_resolve), decisions (get_answer + list_answer_products — the Answer Products catalog), and data (get_artist_status, get_agency, get_person, get_related, get_kculture_calendar, get_korea_rising, get_buy_options) — every response bilingual, provenance-bearing, with a ready-to-cite line.
• Citable by answer engines (AEO/GEO) — per-entity / per-person / per-label citable pages with Schema.org JSON-LD (MusicGroup · TVSeries · Movie · Person · Organization · ItemList · FAQPage · BreadcrumbList), Open Graph, a daily sitemap.xml, a live-generated /llms.txt, and /latest.json.
• YouTube official-channel release/stats (live-state) · LLM romanization at ingest.

Why this exists

Raw Korean API wrappers are a commodity (20+ already exist on GitHub). Our moat is the combination nobody else ships:

• Aggregation of fragmented K-culture / commerce sources
• Verification — Skill Score + provenance, exactly where LLMs confidently hallucinate
• Append-only time-series — a latecomer cannot reconstruct our history
• Behavioral signal — what agents query / buy through us becomes trend data

The customer is the AI agent (consumer); humans / brands / enterprises pay.

Why now — the land-grab window

The compounding assets accrue to early, high-quality entrants: only ~13% of public MCP servers are high-trust, and AI answer engines concentrate citations on content refreshed in the last 1–3 years (Seer Interactive). A verified hub that re-verifies daily compounds a citation lead latecomers can't backfill. We are "picks-and-shovels" — the data agents consume, not a chat wrapper (a category the same market analyses find largely fails to monetize). (An independent 2026 AI-agent opportunity ranking places this exact model at its top — see ROADMAP.md.)

Revenue flywheel (engines ① + ②)

K-culture current-state is the magnet. ① commerce commission + ② trend-intelligence subscription reinforce each other: transactions generate the behavioral signal that becomes the trend product, which improves commerce conversion. See SCOPE.md §3.

The heart: append-only ingestion (component A)

fetch → LLM-extract → cross-verify → bilingual-normalize → append (+ Skill Score)

Overwrite = wrapper. Append timestamped snapshots = an asset.

Bilingual by design

Korean = canonical (provenance anchor). English = distribution layer. Names carry ko / en_official / romanized. See SCOPE.md §5.

Layout

koreaapi/
├── SCOPE.md                 # locked Phase 1 spec
├── llms.txt                 # agent-facing description
├── pyproject.toml
└── src/koreaapi/
    ├── models.py            # bilingual records + Provenance (the data contract)
    ├── skill_score.py       # transparent 0–1 quality score
    ├── pipeline/            # component A: append-only ingestion (the heart)
    │   ├── ingest.py        # fetch→extract→verify→translate→append
    │   ├── store.py         # APPEND-ONLY store (the moat)
    │   └── scheduler.py     # tiered collection cadence
    └── sources/             # source adapters (official APIs first)
        └── base.py

Dev

cd koreaapi
uv sync                      # or: pip install pydantic pytest

# run the offline end-to-end pipeline test (no API keys / network needed)
PYTHONPATH=src python -m pytest tests -q

The append-only ingestion heart (store + ingest + Skill Score + bilingual normalization) is implemented and tested offline via a MockSource. Real source adapters, all with pure fixture-tested parse steps + best-effort live fetch (graceful when egress/keys are absent):

• Wikidata (#1) — bilingual labels via a curated entity→Q-id fast path (each anchor's identity verified, so a contradictory label is rejected, not ingested) + live wbsearchentities. Also pulls the 소속사/label (P264) and discovers labelmates (SPARQL).
• Wikipedia (#2) — independent cross-check; when both agree on the bilingual name the Skill Score clears the single-source cap (the verification moat).
• YouTube Data API (#3.5) — official-channel stats + latest release (live-state event data), identity-guarded; deliberately not a name cross-verifier.
• Circle Chart (#3) — official chart, LLM-extracted with an anti-hallucination grounding guard (entries must appear verbatim in the page HTML). The page is JS-rendered, so the raw chart awaits a data endpoint; the guard ensures it ships nothing over anything false.
• LLM romanization (Haiku) fills romanized at ingest — "cheap AI as collection labor".

Spotify is skipped (its Web API now requires Premium, 2026); a keyless EN-mostly source would only lower the cross-verified scores. See ROADMAP.md for the full log.

Egress note: the live pull needs outbound access to *.wikidata.org. In the web/sandbox environment egress is allowlist-gated — if Wikidata isn't allowlisted the live test skips (HTTP 403 host_not_allowed) while the offline parser tests still cover correctness.

Viewing & managing it (human console)

The product is agent-facing (MCP), but you (human) need a cockpit. There are two faces over one source of truth (the append-only store): the MCP server for agents, and a read-only console for you.

cd koreaapi
PYTHONPATH=src python -m koreaapi.admin seed     # populate koreaapi.db (offline sample)
PYTHONPATH=src python -m koreaapi.admin pull     # LIVE: Wikidata+Wikipedia cross-verified snapshots (+agency)
PYTHONPATH=src python -m koreaapi.admin sweep    # LIVE: discover labelmates from each anchored agency (SPARQL)
PYTHONPATH=src python -m koreaapi.admin youtube  # LIVE: official-channel release snapshots (needs YOUTUBE_API_KEY)
PYTHONPATH=src python -m koreaapi.admin chart    # LIVE: Circle Chart (LLM-extract, grounding-guarded; needs key)
PYTHONPATH=src python -m koreaapi.admin export   # write data/ asset (history + latest.json)
PYTHONPATH=src python -m koreaapi.admin signals  # top behavioral signals (engine 2: what agents query)
PYTHONPATH=src python -m koreaapi.admin stats    # data-quality summary
PYTHONPATH=src python -m koreaapi.admin dump     # print recent snapshots
PYTHONPATH=src python -m koreaapi.admin report   # -> report.html (open in a browser)

# zero-code interactive browse + query + JSON API over the same DB:
pip install datasette && datasette koreaapi.db

Automated collection (cron). .github/workflows/collect.yml runs admin pull + admin export daily (and on manual dispatch) and commits the growing data asset back to the repo: koreaapi/data/snapshots.jsonl (append-only history) + latest.json (current state, crawlable for GEO). It runs on GitHub's runners — open network, so the live pull works there even though the dev sandbox blocks Wikidata egress. Production scales this to Postgres behind the same insert-only contract (see pipeline/store.py); the repo file set is the zero-cost cold-start "database".

Public GEO page. .github/workflows/pages.yml builds report.html from live data and deploys it to GitHub Pages (one-time enable: Settings → Pages → Source: GitHub Actions) — a public, crawlable, JSON-LD-bearing URL so answer engines can surface and cite the verified data.

Watch the headline metric of a verifiable-data business: avg Skill Score, freshness, and source agreement - that is literally watching the moat.

Agent face (MCP server)

The product itself: an MCP server exposing 11 tools, each returning verified, bilingual, provenance-bearing data (with a ready-to-cite line) from the same store the console reads.

cd koreaapi
pip install fastmcp                           # use a venv if system deps clash
PYTHONPATH=src python -m koreaapi.server      # serves over MCP (stdio)

Logic lives in service.py (pure, offline-tested); server.py is the thin MCP binding. Tools register cleanly (verified in an isolated venv).

Install / connect it in your agent: see docs/MCP_INSTALL.md (run command, Claude-Desktop config, and smithery.yaml for the Smithery registry).