GLIN PROFANITY
The Open-Source AI Guardrail. Profanity, PII, Secrets, Soon: Prompt Injection β One Library, One MCP Server, Runs Offline.
---
## π¦ Packages
This monorepo maintains the following packages:
| Package | Version | Description |
|---------|---------|-------------|
| [glin-profanity](https://www.npmjs.com/package/glin-profanity) | [](https://www.npmjs.com/package/glin-profanity) | Core profanity filter for JavaScript/TypeScript |
| [glin-profanity](https://pypi.org/project/glin-profanity/) | [](https://pypi.org/project/glin-profanity/) | Core profanity filter for Python |
| [glin-profanity-mcp](https://www.npmjs.com/package/glin-profanity-mcp) | [](https://www.npmjs.com/package/glin-profanity-mcp) | MCP server for AI assistants (Claude, Cursor, etc.) |
| [openclaw-profanity](https://www.npmjs.com/package/openclaw-profanity) | [](https://www.npmjs.com/package/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 `fΥ½ck` (Cyrillic `Υ½` β `u`). LLM pipelines leak PII and secrets into logs. Prompt injection slips through unguarded inboxes. Today's moderation problem is a guardrail problem β 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** β Python + Llama Community License, requires downloading weights, no Node support, no MCP server.
**vs. ProtectAI llm-guard** β Python-only, heavy transformer dependencies, no edge/browser runtime.
**vs. Azure Content Safety** β paid cloud API, data leaves your infra, rate-limited.
```
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β GLIN PROFANITY v3 β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β β
β Input Text βββΊ Unicode βββΊ Leetspeak βββΊ Dictionary βββΊ ML β
β Normalization Detection Matching Checkβ
β (homoglyphs) (f4ckβfuck) (24 langs) (opt) β
β β
β "fΥ½ck" βββΊ "fuck" βββΊ "fuck" βββΊ MATCH βββΊ β β
β β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
```
---
## 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**
```bash
npm install glin-profanity
```
**Python**
```bash
pip install glin-profanity
```
---
## Quick Start
**JavaScript**
```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**
```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**
```tsx
import { useProfanityChecker } from 'glin-profanity';
function ChatInput() {
const { result, checkText } = useProfanityChecker({
detectLeetspeak: true
});
return (
checkText(e.target.value)} />
{result?.containsProfanity && Clean up your language}
);
}
```
---
## Architecture
```mermaid
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
```javascript
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
```javascript
const filter = new Filter({ normalizeUnicode: true });
filter.isProfane('fΥ½ck'); // true (Armenian 'Υ½' β 'u')
filter.isProfane('shΡt'); // true (Cyrillic 'Ρ' β 'i')
filter.isProfane('Ζuck'); // true (Latin 'Ζ' β 'f')
```
### ML-Powered Detection
```javascript
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](./docs/getting-started.md) | Installation and basic usage |
| [API Reference](./docs/api-reference.md) | Complete API documentation |
| [Framework Examples](./docs/framework-examples.md) | React, Vue, Angular, Express, Next.js |
| [Advanced Features](./docs/advanced-features.md) | Leetspeak, Unicode, ML, caching |
| [ML Guide](./docs/ML-GUIDE.md) | TensorFlow.js integration |
| [Changelog](./CHANGELOG.md) | Version history |
---
## Local Testing Interface
Run the interactive playground locally to test profanity detection:
```bash
# 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`):
```json
{
"mcpServers": {
"glin-profanity": {
"command": "npx",
"args": ["-y", "glin-profanity-mcp"]
}
}
}
```
**Cursor** (`.cursor/mcp.json`):
```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](./packages/mcp/README.md) for setup instructions and examples.
---
## Shipped AI Guardrails
The scanner layer is live. Import from `glin-profanity/scanners`:
```js
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, β¦) + Shannon entropy |
| `PiiScanner` | 27 patterns with Luhn + IBAN mod-97 validation |
| `Vault` | Placeholder-based redact/restore with 4 strategies |
| `scanAll` | Composite scanner β 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](./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](https://glincker.com).
---
## Community
Join the [GLINR Discord](https://discord.gg/Ar5pcaZB99) to ask questions, share feedback, or talk with other users of Glin Profanity. This project has its own forum channel there, `#glin-profanity`.
---
## Contributing
See [CONTRIBUTING.md](./CONTRIBUTING.md) for guidelines. We welcome:
- Bug reports and fixes
- New language dictionaries
- Performance improvements
- Documentation updates
---
## Star History
---
**[Live Demo](https://www.glincker.com/tools/glin-profanity)** Β· **[NPM](https://www.npmjs.com/package/glin-profanity)** Β· **[PyPI](https://pypi.org/project/glin-profanity/)** Β· **[GitHub](https://github.com/GLINCKER/glin-profanity)**
Built by GLINCKER Β· Part of the GLINR ecosystem