content-genie-mcp

Korean-content-creator MCP that gives Claude live Naver/Daum/Google/YouTube/Zum trends, a 100+ Korean-holiday DB, and a viral-score predictor — so you stop tab-hopping between 5 trend dashboards before every post.

English · 한국어

Why content-genie?

Most "trend" MCPs are thin wrappers around Google Trends. If you're publishing for a Korean audience (Naver Blog, YouTube Korea, Instagram KR, TikTok), Google alone misses the signal — Naver Datalab, Daum search, Zum realtime, and KR-region YouTube each surface different keywords at different hours. content-genie consolidates all five sources into one MCP, layers a 100+ Korean-holiday + 14일-day-series + 절기 event database so Claude can plan around 빼빼로데이 and 수능 without being told, and finishes the loop with a viral-score predictor trained on the structural patterns (numbers, brackets, emotional triggers, urgency words) that move CTR in Korean content.

The Phase 4 refactor (v2.12.0, May 2026) made the whole thing safe to depend on: every scraper sits behind a per-source circuit breaker (3 fails → open 5 min) with a stale-cache fallback, so a Naver HTML change no longer takes the whole MCP down — it just degrades that one source and tells the LLM about it via resource://content-genie/sources.

What's new in v2.12.0

• 17 tools, one registry, six small files. The old 4,712-line index.ts is gone — tools live in src/tools/{trends,seo,contentIdeas,viralScoring,competitorAnalysis,koreanEvents}.ts. Adding a tool is a one-liner.
• Scrapers can't take the server down anymore. Per-source circuit breaker + LRU/TTL cache (15 min, 100 entries) + stale fallback. When Naver is degraded, Claude sees source_status: 'stale' and adapts.
• Two new MCP Resources for @-mention. resource://content-genie/korean-events/{year} returns the full event DB rebased to any year; resource://content-genie/sources exposes live breaker + cache state.
• Two MCP Prompts as guided workflows. viral-title chains optimize_title_hashtags → predict_viral_score → generate_ab_test_variants. competitor-analysis does URL analysis + gap-filling ideation.
• 38 tests across 8 suites. Circuit breaker transitions, cache eviction, scraper fallback paths, SSRF guard, every tool group — runs on Node 20 + 22 in CI.

5-minute Quickstart

# 1. Install via Claude Code (recommended)
claude mcp add content-genie -- npx -y content-genie-mcp

# 2. Or add to Claude Desktop's claude_desktop_config.json:

{
  "mcpServers": {
    "content-genie": {
      "command": "npx",
      "args": ["-y", "content-genie-mcp"]
    }
  }
}

3. Restart Claude Desktop (or claude will pick it up on the next run).
4. Try this in Claude:

오늘 한국에서 뜨는 키워드 5개로 유튜브 쇼츠 제목 후보 3개씩 만들고, 각각 바이럴 점수 등급 매겨줘.

5. Expected: Claude calls get_korean_trends (naver+daum+google+youtube+zum merged) → optimize_title_hashtags per keyword → predict_viral_score → returns a ranked S/A/B/C/D grade table with the structural reason ("제목에 숫자 + 호기심 트리거 포함").

Real use cases

1. "I need a Korean blog post topic that won't be obsolete by tomorrow"

Problem: Naver Blog rewards quick takes on the keywords trending right now, but those keywords change every 4 hours. With this MCP: Claude calls get_korean_trends({ platform: 'naver', limit: 20 }) and then generate_content_ideas against the freshest 5 keywords with seasonal: true. The seasonal layer adds upcoming Korean events from the 100+ DB — so the suggestion isn't just "AI 추천 도구" but "AI 추천 도구 + 수능 D-30 콘텐츠 기획" because get_seasonal_content_guide saw 수능 is 4 weeks out. Why it's better than the Naver Datalab UI: Datalab gives you a chart. This MCP gives you 5 titled, hashtag-optimized post drafts grounded in those same trends, in one conversational round-trip.

2. "Will this YouTube Shorts title actually pop?"

Problem: Title CTR is the single biggest lever on Shorts, but you only know after you publish. With this MCP: Paste your draft into Claude, say "use the viral-title prompt". The prompt chains optimize_title_hashtags → predict_viral_score → generate_ab_test_variants and returns an S~D grade plus 5 A/B variants with the structural deltas (added number, added urgency word, shortened to 25 chars). Why it's better than vibes: the scorer reads the same 6 signal classes a Korean copywriter would — emotional triggers (충격|실화|숨겨진), structural elements (numbers/brackets/?/!), urgency, social proof, utility framings (방법|꿀팁|노하우), and length window — and tells you which ones are missing.

3. "I run a small Korean brand, I need a content calendar that respects local holidays"

Problem: Generic calendar tools don't know about 빼빼로데이, 김장철, 수능, or 한가위 연휴, so you end up posting a promo on 현충일. With this MCP: Claude calls create_content_calendar({ months: 3, niche: 'beauty' }), which joins your 3-month window against the 100+ event DB (공휴일 18 + 데이 시리즈 12 + 절기 15 + 상업 이벤트 15 + 시즌/날씨 12 + 입시 10 + 쇼핑 10 + 크리에이터 특화 8) and produces a date-by-date plan with content themes per event. Why it's better than a Notion template: the DB is rebased to any year (resource://content-genie/korean-events/2027 is a single @-mention away), so the calendar stays accurate next year without you maintaining it.

Tools / Resources / Prompts

How it works

                    Claude (or any MCP client)
                              │
                              ▼
                        McpServer (SDK 1.25)
                              │
                  ┌───────────┼───────────┐
                  ▼           ▼           ▼
              17 tools    Resources    Prompts
                  │
                  ▼
            runScraper()  ◄── core/cache.ts (LRU + 15min TTL)
                  │       ◄── core/circuitBreaker.ts (3-fail → open 5min)
                  ▼
         scrapers/{naver,daum,google,youtube,zum}.ts
                  │
                  ▼
            core/security.ts (SSRF + 30s timeout + 5MiB cap)

Three design choices that matter:

1. Per-source circuit breaker. A bad day at Naver doesn't take down 16 other tools — only the trend tools that need Naver, and even those return a stale-cache status: 'stale' envelope so the LLM can adapt.
2. Scrapers never throw. They return a ScrapeResult envelope (ok | stale | unavailable). All error handling is in one place (runScraper), so adding a new source is "write a scraper, return ScrapeResult, done".
3. Event DB rebased on read. The 100+ Korean holiday/event database is stored as month-day rules and rebased to any year on demand — so …/korean-events/2027 works today.

Configuration

HTTP endpoints (when MCP_HTTP_MODE=true):