privacy-guard-proxy

privacy-guard-proxy is a local reverse proxy for Claude Code that masks sensitive data before it leaves your machine.

Email addresses, names, IBANs, API keys, IDs, addresses, and other secrets are replaced with placeholders like [EMAIL_1] or [IBAN_1] on the way to the Anthropic API. Detection runs fully in-process. No external redaction service is involved.

Why use it?

• Keep PII and secrets out of model requests by default
• Run locally with no extra SaaS dependency
• Use it as a drop-in proxy for Claude Code
• Inspect detections, logs, and config in a built-in Web UI
• Expose the same detector engine through a REST API

How it works

Claude Code -> privacy-guard-proxy -> Anthropic API
              | masks PII + secrets |

Before a request is forwarded upstream, the proxy scans user messages, tool payloads, and file contents for sensitive data. Matching values are replaced with stable placeholders. Upstream responses are passed through unchanged.

Quick start

Local binary

cp config.example.json config.json
go build -o privacy-guard-proxy .
./privacy-guard-proxy

export ANTHROPIC_BASE_URL=http://127.0.0.1:9880
claude

By default:

• Proxy: http://127.0.0.1:9880
• Web UI + REST API: http://127.0.0.1:9881

Docker Compose

cp config.example.json config.json
docker compose up -d

The compose setup exposes the same two ports and mounts config.json plus ./logs.

What gets masked?

Web UI

If api_port is enabled, the browser UI is available at http://localhost:9881.

First login:

• Username: admin
• Password: admin

You are forced to change the password on first login.

The UI lets you:

• Manage one or more proxy instances
• Choose detectors and whitelists per proxy
• Enable dry_run mode to log findings without modifying requests
• Manage REST API keys without storing plaintext secrets
• Configure default detectors, default whitelist, and CORS
• Inspect live proxy logs
• Test detector behavior interactively

REST API

The built-in API exposes the same detector engine used by the proxy.

# Anonymize text
curl http://localhost:9881/anonymize \
  -H "Content-Type: application/json" \
  -d '{"text":"Meine IBAN: DE89370400440532013000"}' | jq

# Full scan with findings and placeholder mapping
curl http://localhost:9881/scan \
  -H "Content-Type: application/json" \
  -d '{"text":"foo@example.com, +49 171 1234567"}' | jq

# Use selected detectors for this request only
curl http://localhost:9881/scan \
  -H "Content-Type: application/json" \
  -d '{"text":"...", "detectors":["EMAIL","IBAN","SECRET"]}' | jq

# Health and metrics
curl http://localhost:9881/health
curl http://localhost:9881/metrics

Authentication is optional. If no API keys are configured, requests are accepted without auth.

curl http://localhost:9881/scan \
  -H "X-Api-Key: pgk_..." \
  -H "Content-Type: application/json" \
  -d '{"text":"..."}'

Current limit:

• /scan and /anonymize: 120 requests per minute
• Limit is applied per API key, or per client IP when auth is disabled

Configuration

Start from config.example.json and adjust as needed:

{
  "api_port": 9881,
  "api_keys": [],
  "default_detectors": [],
  "default_whitelist": [],
  "cors_origins": [],
  "ui_password_hash": "",
  "proxies": [
    {
      "type": "claude-code",
      "port": 9880,
      "upstream": "https://api.anthropic.com",
      "privacy_guard": {
        "url": "",
        "api_key": null,
        "detectors": [],
        "whitelist": [],
        "dry_run": false
      }
    }
  ]
}

For /scan and /anonymize, the runtime whitelist is the combination of default_whitelist and any request-level whitelist.

Legacy configs using a top-level []Config array are still accepted.

Supported detectors

Typical use cases

• Protect prompts and file snippets sent from Claude Code
• Offer a local anonymization API for internal tooling
• Run detector tests against sample text before rolling out changes
• Operate in dry_run mode first to see what would be masked

Development

go build -o privacy-guard-proxy .
go test ./...

Project layout:

internal/
├── api/        # Web UI + REST API
├── detector/   # PII and secret detection engine
└── proxy/      # Reverse proxy and masking pipeline
main.go         # Application entry point
config.json     # Runtime configuration

Frontend assets are vendored under internal/api/static/vendor/, so the Web UI does not depend on external CDNs at runtime.