AI Privacy Gateway intercepts every API call to LLMs — detecting and masking phone numbers, emails, ID cards, bank cards, and 14 entity types
{
"model": "gpt-4",
"messages": [{
"role": "user",
"content": "My phone is 13812345678,
email is alice@example.com,
and my ID is 110101199001011234."
}]
}Works with ChatGPT, Claude, DeepSeek, Cursor, Kimi — any OpenAI-compatible API.
Every time you paste code, patient records, or customer data into an AI prompt, that information leaves your control. Here is what is at stake.
Phone numbers, emails, addresses, and bank cards fly through third-party servers every time you use an AI tool.
LLMs memorize patterns. Your data could resurface in another user's session — a documented risk in production AI systems.
Sending PII to third-party AI APIs without masking violates GDPR, CCPA, PIPL, and HIPAA. Fines can reach millions.
Point your API base URL to the gateway. Everything else is automatic.
Gateway captures every request headed to your AI provider
14 entity types identified via regex + NER, replaced with safe placeholders
Masked data sent to LLM; response placeholders restored before returning to you
# Unprotected — direct to AI
client = OpenAI(
api_key="sk-...",
base_url="https://api.openai.com/v1"
)# Protected — via gateway
client = OpenAI(
api_key="sk-...",
base_url="http://localhost:9999/v1"
)Built for teams that ship fast and care about data security. Every feature ships in the free Lite edition.
Regex + NER dual-engine detects 14 entity types: phone, email, ID card, bank card, name, address, license plate, IP, URL, date, amount, postal code, organization, and custom keywords. Zero configuration.
Full SSE proxy with real-time masking and unmasking. Zero added perceptible latency.
Encrypted local storage. Mapping data never leaves your machine.
Single binary, Docker, or pip. Download and run. No dependencies.
Request enters gateway → regex scan → NER inference → placeholder replacement → forward to LLM. Response reverses the pipeline. End-to-end under 1ms overhead.
Regex + keyword patterns for domain-specific PII. Flexible and extensible.
Real-time stats: masking count, entity distribution, mapping records, and throughput.
Free forever. Personal use, commercial use, self-hosted. No restrictions.
P99 latency under 2ms. Single instance handles 10,000+ QPS. Zero data leaves your network.
Drop-in proxy for any OpenAI-compatible API. No SDK changes, no vendor lock-in.
Deploy in 30 seconds. No credit card. No account. No data ever leaves your machine.