Masker

PII redaction API for LLMs. Remove personal data before sending to ChatGPT, Claude, or any AI.

Open the project site · Run locally

Docs · Examples · API quickstart

What it does

Input: "Contact John Doe at john@example.com"
Output: "Contact <PERSON> at <EMAIL>"

Masker detects and redacts:

• EMAIL — email addresses
• PHONE — phone numbers
• CARD — credit card numbers
• PERSON — person names (AI-powered)
• SENSITIVE_WORD — custom words and phrases configured by the operator

Quick Start

docker run -p 8000:8000 ghcr.io/kikuai-lab/masker

curl -X POST http://localhost:8000/v1/redact \
-H "Content-Type: application/json" \
-d '{"text": "Email me at john@example.com", "mode": "placeholder"}'

LLM Proxy (OpenAI-compatible)

Drop-in replacement for /v1/chat/completions that automatically redacts PII:

importopenaiclient=openai.OpenAI(
base_url="http://localhost:8000/v1",
api_key="your-openai-key",
default_headers={"X-Api-Key": "your-masker-key"}
)
response=client.chat.completions.create(
model="gpt-4",
messages=[{"role": "user", "content": "My email is john@example.com"}]
)
# PII redacted before reaching OpenAI

Features

• Text & JSON — process plain text or nested JSON
• Policy system — configure mask/drop/placeholder per PII type
• Audit logging — JSONL logs without storing raw text
• Fail modes — closed (block) or open (pass-through)
• Multi-tenant — API key authentication
• Multi-language — English & Russian NER

Installation

git clone https://github.com/KikuAI-Lab/masker.git
cd masker
cp .env.example .env
docker-compose up -d

Configuration

MASKER_API_KEYS=sk-key1:tenant1,sk-key2:tenant2
MASKER_UPSTREAM_URL=https://api.openai.com/v1/chat/completions
MASKER_DEFAULT_FAIL_MODE=closed
MASKER_SENSITIVE_WORDS=secret,confidential,internal project

MASKER_SENSITIVE_WORDS is a comma-separated, case-insensitive list.

See .env.example for all options.

Documentation

📖 Wiki — full documentation

• Quick Start
• API Reference
• LLM Proxy
• Privacy Policy

API

Endpoint	Description
POST /v1/redact	Redact PII from text or JSON
POST /v1/chat/completions	OpenAI-compatible proxy
GET /health	Health check
GET /metrics	Prometheus metrics

Security & Architecture

Stateless by Design. Masker is a firewall, not a storage bucket.

• ✅ No raw payload storage by design — Payloads are processed in RAM and are not intentionally persisted.
• ✅ Local PII detection path — Redaction logic does not require third-party AI calls.
• ✅ Self-hostable boundary — The redaction service can run without internet access; proxy mode still calls the configured upstream provider.

📖 Read about our Stateless Architecture to understand the intended privacy boundary and deployment assumptions.

Follow the work

Project notes and new tools: Telegram · LinkedIn · KikuAI

License

AGPL-3.0 — KikuAI OÜ

_{Follow new projects and updates from @kiku-jw.}