Kadlint MCP server

kadlint

Korean ad-compliance linter for cosmetics and health-food marketing copy — detects prohibited claims with the legal basis, severity, and safer alternatives.

한국어 README

Korean law tightly regulates what cosmetics and health-functional-food ads may claim. A face cream may not say it "treats acne" (that is a drug claim under the Cosmetics Act), a probiotic may not promise to "boost immunity" unless that exact functionality was approved, and "No.1" / "100% natural" claims require objective substantiation under the Fair Labeling and Advertising Act. Korea's Ministry of Food and Drug Safety (MFDS) keeps detecting thousands of unlawful online ads in these categories every year — in 2025 its routine sweeps again flagged large batches of cosmetics and health-food ads for drug-like or unsubstantiated claims, and violations can mean ad bans, sales suspensions, and fines.

kadlint lints Korean marketing copy against a curated dictionary of 299 rules (183 cosmetics, 51 health-food, 65 common) distilled from:

• 화장품법 (Cosmetics Act) and its enforcement rules (별표5)
• 건강기능식품법 / 식품표시광고법 (Health Functional Foods & Food Labeling/Advertising Acts)
• 표시광고법 (Fair Labeling and Advertising Act)
• MFDS advertising guidelines and enforcement precedents

Every finding comes with the violation category, the legal basis citation, a severity (high / medium / low), an explanation, and safer alternative phrasings.

Install

npm install -g kadlint     # CLI
npm install kadlint        # library

Requires Node.js >= 18.

CLI

kadlint check ad-copy.txt
kadlint check ./landing-pages --format json --min-severity high
kadlint check - < draft.md                 # stdin
kadlint check banner.html --domain cosmetics,common

ad-copy.txt:1:1 [HIGH] '여드름 치료' — 의약품 오인 표현 (cos-113)
    근거: 화장품법 제13조제1항제1호, 식약처 「화장품 표시·광고 관리 지침」
    대체 표현: 피지·각질 케어로 매끈한 피부결 / (실증 보유 시) 여드름성 피부에 사용 적합

✖ 1건 발견 (high 1 · medium 0 · low 0) — 파일 1개, 규칙 299개 검사

• Supported inputs: .txt, .md, .html (tags stripped, offsets preserved), directories (recursive), - (stdin).
• Exit codes: 0 clean, 1 findings at/above --min-severity, 2 usage error.
• Literal rules match through inserted spaces, middle dots, and hyphens (살 균, 살·균, 살-균), a common evasion tactic; disable with --no-normalize.

API

import { lint, lintFile } from 'kadlint';

const result = lint('여드름 치료에 탁월한 크림', { domains: ['cosmetics'] });
for (const f of result.findings) {
  console.log(f.ruleId, f.severity, f.match, f.legalBasis, f.alternatives);
}

const fileResult = lintFile('landing.html', { minSeverity: 'medium' });

Options: domains (cosmetics | health_food | common), minSeverity, normalize (default true).

When one of a rule's exceptions phrases appears within 80 characters of a match (e.g. an approved functional claim used verbatim), the finding is downgraded to low with an explanatory note.

MCP server

kadlint-mcp exposes a checkadcopy { text, domain? } tool over stdio so LLM assistants can lint copy while drafting it.

Claude Desktop (claudedesktopconfig.json):

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

If installed globally, "command": "kadlint-mcp" suffices:

{
  "mcpServers": {
    "kadlint": { "command": "kadlint-mcp" }
  }
}

Claude Code:

claude mcp add kadlint -- kadlint-mcp

or in .mcp.json:

{
  "mcpServers": {
    "kadlint": {
      "command": "kadlint-mcp",
      "args": []
    }
  }
}

Playground

Shortened. The full README is on GitHub.