Detect, clean, and analyze profanity in text via profanease (npx) — l33t speak, 25 languages, 7 categories, 4 replacement styles

profanease v2.0.3 — Profanity Detection & Content Moderation MCP Server

Install & run: npm install profanease → node node_modules/profanease/dist/mcp-server.js (stdio) Zero runtime dependencies. Pure local — no external API calls, no auth.

3 Tools

Key Parameters

• `language`: "en" (default, English only) or "all" (25 languages including French, German, Spanish, etc.)
• `normalize`: "none" | "basic" | "moderate" (default) | "aggressive" — controls l33t speak / homoglyph detection
• `categories`: filter to specific categories: profanity, sexual, slur, insult, religious, drugs, violence
• `replacement`: "asterisk" (→ ****), "grawlix" (→ @#$%), "word" (→ [censored]), "full" (repeated placeholder)
• `custom_words` + `custom_only`: brand-specific / domain-specific filtering without built-in lists

Verified Execution Trace (12 calls, 100% success, p50=1.5ms)

1. check — clean text: {text: "Hello, this is a perfectly clean message."} → {isProfane: false} (5ms)
2. check — profanity: {text: "What the hell is going on here?"} → {isProfane: true} (2ms)
3. check — l33t speak (moderate): {text: "You are such a $h1t person", normalize: "moderate"} → {isProfane: true} (1ms) ✅ caught
4. check — l33t speak (none): {text: "You are such a $h1t person", normalize: "none"} → {isProfane: false} (1ms) ✅ correctly missed
5. check — violence category only: {text: "I will kill you, you idiot", categories: ["violence"]} → {isProfane: false} (2ms) ⚠️ "kill" not in violence list
6. check — custom words only: {text: "The competitor product is terrible", custom_words: ["competitor","terrible"], custom_only: true} → {isProfane: true} (1ms)
7. clean — asterisk: {text: "What the hell, you damn fool!"} → {cleaned: "What the ****, you **** fool!"} (1ms)
8. clean — grawlix: same text, replacement: "grawlix" → {cleaned: "What the @#$%, you @#$% fool!"} (2ms)
9. clean — word: {text: "That is total crap and BS", replacement: "word"} → {cleaned: "That is total [censored] and BS"} (1ms)
10. analyze — mixed: {text: "Go to hell you stupid idiot, this is crap"} → 3 matches (hell, stupid, crap), severity "moderate" (1ms) ⚠️ "idiot" NOT detected
11. analyze — multi-language: {text: "This is merde and scheisse", language: "all"} → 2 matches (merde, scheisse), severity "mild" (2ms) ✅ French+German
12. analyze — aggressive normalize: {text: "You are an a$$hole", normalize: "aggressive"} → 1 match (a$$hole→asshole), severity "mild" (7ms) ✅

Critical Gotchas

1. `categories` filter exists but match results always show `categories: []` — the per-match category classification appears non-functional even though the filter param works for narrowing scope
2. "idiot" is NOT in the built-in word list despite being a common insult — vocabulary is conservative
3. "BS" abbreviations NOT detected — only full words are matched
4. "kill" is NOT in the violence category — the violence category has a narrow wordlist
5. Aggressive normalize produces odd cleaned output — "a$$hole" cleans to "*$$****" (original special chars partially preserved in replacement)
6. severity levels are global (mild/moderate/severe), not per-match — computed across all matches
7. First call ~5ms (JIT/init), subsequent calls ~1-2ms
8. Multi-language `"all"` mode may produce false positives on short words that happen to m