I built an open source proxy that blocks PII and prompt injection before any LLM call

JEEVAN NAGARAJ — Tue, 13 Oct 2026 14:57:06 +0000

The problem

Every time your application calls Claude or GPT-4o, your users' sensitive data travels to an external server. Names, emails, phone numbers, NHS numbers, financial data. Most developers have no visibility into what is actually leaving their system.

What I built

Jnvoy is a drop-in proxy. Change one URL in your code and every LLM call is automatically protected.

Before Jnvoy:
Your app sends raw user data directly to Claude or GPT-4o.

After Jnvoy:
Every request is scanned locally, PII is replaced with reversible tokens, the clean request goes to the LLM, and original values are restored in the response.

What it does

Microsoft Presidio PII detection with 20+ entity types
UK-specific recognisers: NHS numbers, NI numbers, UK postcodes, sort codes
Prompt injection detection: blocks jailbreaks, instruction overrides, system prompt extraction
MCP security: detects tool poisoning and secret leakage in tool responses
PostgreSQL audit logs for FCA and GDPR compliance
One click FCA and EU AI Act compliance report PDF
Semantic caching saves 20 to 40 percent of LLM costs
Multi-tenant API keys with complete data isolation
Slack connector protecting AI bot messages
Docker stack in one command
164 automated tests passing

One URL change

# Before
import anthropic
client = anthropic.Anthropic(api_key="sk-ant-...")
response = client.messages.create(model="claude-haiku-4-5", messages=[{"role": "user", "content": user_message}])

# After
import httpx
response = httpx.post("https://jnvoy.onrender.com/v1/proxy", json={"model": "claude-haiku-4-5", "provider": "anthropic", "messages": [{"role": "user", "content": user_message}]})

Try it

curl "https://jnvoy.onrender.com/v1/scan?text=My NI number is AB123456C and email is john@example.com"

Live: https://jnvoy.com
GitHub: https://github.com/Jeevan2191999/jnvoy