Open-source PII masking proxy for ChatGPT, Claude, Cursor, DeepSeek, and any LLM API. Install a firewall for your AI data in 30 seconds. [MIT Licensed]
v2.0.0 — A high-performance reverse proxy that automatically detects and masks sensitive data (phone numbers, ID cards, emails, bank cards, names, locations, API keys, and 14+ entity types) in AI API requests and responses. Protects PII before it leaves your machine. New in v2.0: AES-256-GCM vault encryption, multi-upstream load balancer, pub/sub audit bus, browser extension SDK, Windows/macOS installers.
Website · Online Demo · Documentation · Releases
Quick Start · Why · Architecture · Comparison · Entities · Deployment
| Concern | Without Gateway | With Gateway |
|---|---|---|
| PII in prompts | Phone numbers, emails, ID cards sent to AI APIs in plain text | Auto-detected and masked as [PHONE_1], [EMAIL_1] before leaving your machine |
| API key leaks | Source code containing API keys sent to third-party servers | Keys detected and replaced with placeholders |
| GDPR / HIPAA compliance | PII transferred without safeguards | PHI masked before transmission; data minimization enforced |
| Model training data | Your data could be used for training | Masked data is anonymous; original PII never stored |
| Latency impact | N/A | Under 1ms average added latency |
Zero configuration. No code changes. Works with any OpenAI-compatible API.
Windows:
# 1. Double-click install.bat — or run from terminal:
install.bat
# 2. Double-click start.bat — or:
start.batmacOS / Linux:
# 1. Run the installer:
chmod +x install.sh && ./install.sh
# 2. Start the gateway:
./start.shThe installer checks your Python version, creates a virtual environment, installs dependencies, generates a secure .env config with a random admin password, and creates a desktop shortcut (Windows). The start script auto-detects your environment, checks port availability, and launches the gateway.
Open http://localhost:9999 to access the admin dashboard.
Non-interactive mode for CI/CD:
python start.py --non-interactive
docker pull ghcr.io/gunxueqiu6/ai-privacy-gateway:lite
docker run -d --name ai-privacy-gw -p 9999:9999 \
ghcr.io/gunxueqiu6/ai-privacy-gateway:lite
# Point your AI client to http://localhost:9999 — done.pip install -r requirements.txt
python main.pySet your AI client's API endpoint to http://localhost:9999:
from openai import OpenAI
client = OpenAI(
base_url="http://localhost:9999/v1",
api_key="your-api-key"
)Settings → API Key → Base URL → http://localhost:9999
Works transparently with any AI tool that supports custom API endpoints.
flowchart LR
A[AI Client<br/>Cursor/Claude Code/ChatGPT] -->|raw request| B[Privacy Gateway<br/>:9999]
B -->|masked request| C[AI API Provider<br/>OpenAI/DeepSeek/Anthropic]
B --> D[(Local SQLite<br/>Encrypted Vault)]
B --> E[Audit Log<br/>JSON/Syslog]
subgraph "Your Machine / VPC"
A
B
D
E
end
subgraph "External"
C
end
Request Flow:
- Your AI client sends a prompt containing sensitive data
- Gateway intercepts, detects 14+ entity types via regex engine (< 1ms)
- PII replaced with typed placeholders:
[PHONE_abc123],[EMAIL_xyz789] - Masked request forwarded to target AI API — provider never sees raw PII
- AI response returned; original values restored from encrypted vault if needed
| Feature | AI Privacy Gateway | LLM Guard | PasteGuard | Nightfall AI | Private AI |
|---|---|---|---|---|---|
| License | MIT | MIT | MIT | Commercial | Commercial |
| Deployment | Docker / pip | pip | Browser Extension | Cloud API | SDK / Cloud |
| Setup time | 30s | 5 min | 1 min | 1 hour+ | 1 hour+ |
| PII types | 14+ | 10+ | 8+ | 30+ | 50+ |
| Latency | < 1ms | ~5ms | < 0.5ms | ~50ms | ~100ms |
| Offline / on-prem | Yes | Yes | Browser-only | No | Partial |
| Audit logging | Yes | Limited | No | Yes | Yes |
| Streaming (SSE) | Yes | No | N/A | No | No |
| HTTPS MITM | Yes | Yes (proxy) | Browser-only | API-based | API-based |
| Cost | Free | Free | Free | $$$$ | $$$ |
- You need zero external dependencies — fully local, no cloud services
- You need streaming support (SSE/real-time masking)
- You want 30-second Docker deployment with no configuration
- You need both transparent proxy AND API mask/restore endpoints
- You're using AI coding tools (Cursor, Claude Code, Copilot) that send code to APIs
| Type | Pattern | Example |
|---|---|---|
| Phone | 1[3-9]\d{9} | 13812345678 |
| ID Card | 18 digits | 110101199001011234 |
| Standard format | user@example.com | |
| Bank Card | 16-19 digits (Luhn check) | 6222021234567890123 |
| Person Name | Chinese + English names | 张三 |
| Location | Cities, districts, provinces | 北京市海淀区 |
| Organization | Company names | 北京科技有限公司 |
| Plate Number | Chinese format | 京A12345 |
| IP Address | IPv4 / IPv6 | 192.168.1.100 |
| URL | HTTP/HTTPS | https://example.com |
| Date | Various formats | 2024年1月15日 |
| Amount | Currency values | ¥999.99 |
| Postcode | 6 digits | 100080 |
| API Key | 20+ known formats | sk-abc... / AKIA... / ghp_... |
| Custom | User-defined regex | Passport numbers, SSN, etc. |
docker-compose up -dapiVersion: v1
kind: Pod
spec:
containers:
- name: app
image: my-app
env:
- name: OPENAI_BASE_URL
value: http://localhost:9999/v1
- name: privacy-proxy
image: ghcr.io/gunxueqiu6/ai-privacy-gateway:lite
ports:
- containerPort: 9999[Unit]
Description=AI Privacy Gateway
After=network.target
[Service]
Type=simple
User=privacygw
WorkingDirectory=/opt/privacy-gateway
ExecStart=/usr/bin/python3 main.py
Restart=always
RestartSec=5
[Install]
WantedBy=multi-user.targetDownload PrivacyGateway.exe from Releases and double-click to run.
Download from Releases, chmod +x PrivacyGateway, and run ./PrivacyGateway.
| Variable | Default | Description |
|---|---|---|
TARGET_LLM |
https://api.openai.com | Target AI API endpoint |
LISTEN_PORT |
9999 | Gateway listen port |
DB_PATH |
./vault_data/privacy_vault.db | SQLite database path |
ADMIN_PASSWORD |
(auto-generated) | Admin dashboard password |
JWT_SECRET |
(auto-generated) | JWT signing secret |
VAULT_ENCRYPT_KEY |
(auto-generated) | AES-256-GCM vault encryption key |
UPSTREAM_LLM_URLS |
(empty) | Comma-separated upstream URLs for load balancing |
UPSTREAM_LB_STRATEGY |
round_robin | Load balancer strategy: round_robin / random / least_connections |
MAPPING_TTL |
259200 (72h) | Mapping TTL in seconds (0 = delete on request complete) |
STATELESS_MODE |
0 | Set to 1 for in-memory only (no disk writes) |
Open http://localhost:9999 and login with your admin password to:
- View real-time interception statistics and trend charts
- Manage custom sensitive words (add, test, delete)
- Check system health and version information
- Browse supported entity types
# Mask text
curl -X POST http://localhost:9999/api/mask \
-H "Content-Type: application/json" \
-d '{"text": "张三住在北京市，电话13812345678"}'
# Restore text
curl -X POST http://localhost:9999/api/restore \
-H "Content-Type: application/json" \
-d '{"text": "[PII_PER_00000001]住在[PII_LOC_00000001]，电话[PII_PHONE_00000001]", "mappings": {...}}'
# Batch mask
curl -X POST http://localhost:9999/api/mask/batch \
-H "Content-Type: application/json" \
-d '{"texts": ["text1", "text2", "text3"]}'ai-privacy-gateway/
├── config.py # Configuration management
├── mask_engine.py # Regex masking engine (14+ types)
├── ner_engine.py # NER entity recognition
├── stream_buffer.py # SSE streaming buffer
├── gateway_core.py # HTTP proxy core
├── database.py # SQLite encrypted vault
├── main.py # FastAPI entry point
├── routers/ # Route modules
│ ├── proxy.py # Core proxy routes
│ ├── api.py # Mask/restore API
│ ├── admin.py # Admin dashboard
│ └── auth.py # Auth endpoints
├── load_balancer.py # Multi-upstream load balancer
├── audit.py # Pub/sub audit event bus
├── vault_crypto.py # AES-256-GCM vault encryption
├── static/ # Admin dashboard UI
├── sdk/ # Client SDKs
│ ├── browser-extension/ # Chrome/Edge extension
│ ├── js/ # JavaScript/TypeScript SDK
│ └── flutter/ # Flutter/Dart SDK
├── tests/ # Test suite
└── website-astro/ # Public website (Astro)
git clone https://github.com/gunxueqiu6/ai-privacy-gateway
cd ai-privacy-gateway
pip install -r requirements.txt
# Run tests
pytest tests/ -v
# Run server
python main.pyMIT License. See LICENSE for details.