GLIN PROFANITY

The Open-Source AI Guardrail. Profanity, PII, Secrets, Soon: Prompt Injection — One Library, One MCP Server, Runs Offline.

---

📦 Packages

This monorepo maintains the following packages:

Package	Version	Description
glin-profanity		Core profanity filter for JavaScript/TypeScript
glin-profanity		Core profanity filter for Python
glin-profanity-mcp		MCP server for AI assistants (Claude, Cursor, etc.)
openclaw-profanity		Plugin for OpenClaw/Moltbot AI agents

---

Why Glin Profanity?

Modern AI applications need more than a word list. Users evade filters with f4ck, sh1t, and fսck (Cyrillic ս → u). LLM pipelines leak PII and secrets into logs. Prompt injection slips through unguarded inboxes. Today's moderation problem is a guardrail problem — and most solutions leave you choosing between a Python-only library, a Llama-licensed model, or a paid cloud API.

Glin Profanity is the MIT-licensed, Node-native, MCP-first answer. It runs entirely offline, ships a 12 KB core bundle with no mandatory cloud calls, integrates with Claude/Cursor/Windsurf via 24 MCP tools out of the box, and covers 24 languages with leetspeak and Unicode homoglyph evasion detection built in. Prompt-injection, PII, and secrets scanning are all shipped today.

vs. Meta PurpleLlama — Python + Llama Community License, requires downloading weights, no Node support, no MCP server. vs. ProtectAI llm-guard — Python-only, heavy transformer dependencies, no edge/browser runtime. vs. Azure Content Safety — paid cloud API, data leaves your infra, rate-limited.

┌─────────────────────────────────────────────────────────────────────────────┐
│                           GLIN PROFANITY v3                                 │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│   Input Text ──►  Unicode       ──►  Leetspeak    ──►  Dictionary  ──► ML  │
│                   Normalization      Detection         Matching        Check│
│                   (homoglyphs)       (f4ck→fuck)       (24 langs)     (opt) │
│                                                                             │
│   "fսck"     ──►  "fuck"        ──►  "fuck"       ──►  MATCH       ──► ✓   │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

---

Performance Benchmarks

From the CI shootout gate (benchmarks/shootout/results.md), Node.js v22, 20-input torture-set batches:

Library	ops/sec	F1 (accuracy)	False-Positive Rate
glin-profanity	990	80.6%	0.0%
obscenity	5,112	79.5%	5.9%
bad-words	241	54.2%	0.0%
leo-profanity	338,407	34.6%	0.0%
@2toad/profanity	839,796	56.7%	0.0%

glin-profanity trades raw throughput for zero false positives and the highest F1 in the field. See benchmarks/shootout/results.md for the full per-category breakdown.

---

Feature Comparison

Feature	glin-profanity	obscenity	Detoxify	PurpleLlama	llm-guard
MIT license	Yes	Yes	Apache-2.0	Llama Community	Apache-2.0
Node-native	Yes	Yes	No	No	No
Python package	Yes	No	Yes	Yes	Yes
MCP server (24 tools)	Yes	No	No	No	No
Runs fully offline	Yes	Yes	Yes	Yes (needs weights)	Yes
Leetspeak detection	Yes	Partial	No	No	No
Unicode homoglyph detection	Yes	No	No	No	No
Multi-language support	24 languages	English only	6 languages	English only	English only
ML toxicity detection	Yes (TensorFlow.js, opt-in)	No	Yes (PyTorch)	Yes (Llama)	Yes (transformers)
Edge / browser runtime	Yes	Yes	No	No	No
Bundle size (core, minified)	12 KB	6 KB	N/A	N/A	N/A
Prompt-injection detection	Yes (shipped)	No	No	Yes	Yes
PII / secrets scanning	Yes (shipped)	No	No	No	Yes

---

Installation

JavaScript/TypeScript

npm install glin-profanity

Python

pip install glin-profanity

---

Quick Start

JavaScript

import { checkProfanity, Filter } from 'glin-profanity';

// Simple check
const result = checkProfanity("This is f4ck1ng bad", {
  detectLeetspeak: true,
  languages: ['english']
});

result.containsProfanity  // true
result.profaneWords       // ['fucking']

// With replacement
const filter = new Filter({
  replaceWith: '***',
  detectLeetspeak: true
});
filter.checkProfanity("sh1t happens").processedText  // "*** happens"

Python

from glin_profanity import Filter

filter = Filter({"languages": ["english"], "replace_with": "***"})

filter.is_profane("damn this")           # True
filter.check_profanity("damn this")      # Full result object

React

import { useProfanityChecker } from 'glin-profanity';

function ChatInput() {
  const { result, checkText } = useProfanityChecker({
    detectLeetspeak: true
  });

  return (
    <input onChange={(e) => checkText(e.target.value)} />
    {result?.containsProfanity && <span>Clean up your language</span>}
  );
}

---

Architecture

flowchart LR
    subgraph Input
        A[Raw Text]
    end

    subgraph Processing
        B[Unicode Normalizer]
        C[Leetspeak Decoder]
        D[Word Tokenizer]
    end

    subgraph Detection
        E[Dictionary Matcher]
        F[Fuzzy Matcher]
        G[ML Toxicity Model]
    end

    subgraph Output
        H[Result Object]
    end

    A --> B --> C --> D
    D --> E --> H
    D --> F --> H
    D -.->|Optional| G -.-> H

---

Detection Capabilities

Leetspeak Detection

const filter = new Filter({
  detectLeetspeak: true,
  leetspeakLevel: 'aggressive'  // basic | moderate | aggressive
});

filter.isProfane('f4ck');     // true
filter.isProfane('5h1t');     // true
filter.isProfane('@$$');      // true
filter.isProfane('ph.u" "ck'); // true (aggressive mode)

Unicode Homoglyph Detection

const filter = new Filter({ normalizeUnicode: true });

filter.isProfane('fսck');   // true (Armenian 'ս' → 'u')
filter.isProfane('shіt');   // true (Cyrillic 'і' → 'i')
filter.isProfane('ƒuck');   // true (Latin 'ƒ' → 'f')

ML-Powered Detection

import { loadToxicityModel, checkToxicity } from 'glin-profanity/ml';

await loadToxicityModel({ threshold: 0.9 });

const result = await checkToxicity("You're the worst player ever");
// { toxic: true, categories: { toxicity: 0.92, insult: 0.87, ... } }

---

Supported Languages

24 languages with curated dictionaries:

Arabic	Chinese	Czech	Danish
Dutch	English	Esperanto	Finnish
French	German	Hindi	Hungarian
Italian	Japanese	Korean	Norwegian
Persian	Polish	Portuguese	Russian
Spanish	Swedish	Thai	Turkish

---

Documentation

Document	Description
Getting Started	Installation and basic usage
API Reference	Complete API documentation
Framework Examples	React, Vue, Angular, Express, Next.js
Advanced Features	Leetspeak, Unicode, ML, caching
ML Guide	TensorFlow.js integration
Changelog	Version history

---

Local Testing Interface

Run the interactive playground locally to test profanity detection:

# Clone the repo
git clone https://github.com/GLINCKER/glin-profanity.git
cd glin-profanity/packages/js

# Install dependencies
npm install

# Start the local testing server
npm run dev:playground

Open http://localhost:4000 to access the testing interface with:

• Real-time profanity detection
• Toggle leetspeak, Unicode normalization, ML detection
• Multi-language selection
• Visual results with severity indicators

---

Use Cases

Application	How Glin Profanity Helps
Chat platforms	Real-time message filtering with React hook
Gaming	Detect obfuscated profanity in player names/chat
Social media	Scale moderation with ML-powered detection
Education	Maintain safe learning environments
Enterprise	Filter internal communications
AI/ML pipelines	Clean training data before model ingestion

---

MCP Server for AI Assistants

Glin Profanity includes an MCP (Model Context Protocol) server that enables AI assistants like Claude Desktop, Cursor, Windsurf, and other MCP-compatible tools to use profanity detection as a native tool.

Quick Setup

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

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

Cursor (.cursor/mcp.json):

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

Available Tools (24)

Tool	Description
check_profanity	Check text for profanity with detailed results
censor_text	Censor profanity with configurable replacement
analyze_context	Context-aware analysis with domain whitelists
batch_check	Check multiple texts in one operation
validate_content	Content validation with safety scoring (0-100)
detect_obfuscation	Detect leetspeak and Unicode tricks
get_supported_languages	List all 24 supported languages
explain_match	Explain why text was flagged with reasoning
suggest_alternatives	Suggest clean alternatives for profane content
analyze_corpus	Analyze up to 500 texts for moderation stats
compare_strictness	Compare results across strictness levels
create_regex_pattern	Generate regex patterns for custom detection
track_user_message	Track user messages for repeat offender detection
get_user_profile	Get moderation profile for a specific user
get_high_risk_users	List users with high violation rates
reset_user_profile	Reset a user's moderation history
stream_check	Real-time streaming profanity check
stream_batch	Stream multiple texts with live results
get_stream_stats	Get streaming session statistics
check_prompt_injection	Scan text for prompt injection attacks (rule-based, 50 patterns)
scan_secrets	Detect leaked API keys, tokens, and credentials (110 patterns + entropy)
scan_pii	Detect PII: email, phone, SSN, credit card, IBAN, passport, and more
redact_pii	Redact PII into reversible vault-backed placeholders
restore_pii	Restore PII placeholders to original values via vault session

Plus 5 workflow prompts and 5 reference resources for guided AI interactions.

Example Prompts for AI Assistants

"Check this user comment for profanity using glin-profanity"
"Validate this blog post content with high strictness"
"Batch check these 50 messages for any inappropriate content"
"Analyze this medical text with the medical domain context"

See the full MCP documentation for setup instructions and examples.

---

Shipped AI Guardrails

The scanner layer is live. Import from glin-profanity/scanners:

import { PromptInjectionScanner, SecretsScanner, PiiScanner, Vault, scanAll } from 'glin-profanity/scanners';

Scanner	Coverage
PromptInjectionScanner	50 patterns across 6 attack categories
SecretsScanner	110 patterns (AWS, GCP, Azure, GitHub, Stripe, OpenAI, Anthropic, …) + Shannon entropy
PiiScanner	27 patterns with Luhn + IBAN mod-97 validation
Vault	Placeholder-based redact/restore with 4 strategies
scanAll	Composite scanner — runs all of the above in one call

Coming in 2026

The following capabilities are on the active roadmap.

Feature	ETA	Notes
glincker/glin-guard-small on HF Hub	Q3 2026	Our own distilled toxicity model, MIT weights, designed for edge inference
AI-slop detection	Q3 2026	Pattern-based detector for generic AI-generated prose
Bluesky Ozone labeler adapter	Q4 2026	Drop-in labeler for AT Protocol moderation pipelines
Compliance presets	Q4 2026	Pre-tuned configs for UK OSA, EU DSA, and COPPA requirements

See ROADMAP.md for the full issue backlog and contribution opportunities.

---

License

MIT License - free for personal and commercial use.

Enterprise licensing with SLA and support available from GLINCKER.

---

Contributing

See CONTRIBUTING.md for guidelines. We welcome:

• Bug reports and fixes
• New language dictionaries
• Performance improvements
• Documentation updates

---

Star History

---