Use Claude & ChatGPT Without Leaking PII

The Developer's Dilemma

You are debugging a production issue. The stack trace has customer email addresses in it. The fastest fix? Paste it into Claude and ask for help.

But that data is now:

• Stored in Anthropic's systems
• Possibly used for model training, depending on your plan
• Visible to anyone who can access your chat history

77% of developers paste sensitive data into AI tools. Most do not notice the problem until later.

How the Proxy Fixes This

The Model Context Protocol (MCP) lets a server sit between you and any AI tool. The anonym.legal MCP server uses that position to strip PII before your text hits any model.

It works in four steps:

1. You type a prompt as normal
2. The proxy catches it before sending
3. PII is found and swapped for reversible tokens
4. The AI only sees clean, anonymized text

The AI's reply comes back with real values restored. Your workflow does not change.

Setting Up the Integration

What You Need

• Node.js 18 or later
• Claude Desktop, Cursor, or VS Code with the Claude extension
• An anonym.legal API key — get one free

Step 1: Get Your API Key

1. Sign up at anonym.legal/auth/signup
2. Go to Settings → API Tokens
3. Create a new token
4. Copy it — you only see it once

Step 2: Configure Claude Desktop

Edit the config file for your OS:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json
Linux: ~/.config/Claude/claude_desktop_config.json

Add the anonym.legal server:

{
  "mcpServers": {
    "anonym-legal": {
      "command": "npx",
      "args": ["-y", "@anonym-legal/mcp-server"],
      "env": {
        "ANONYM_API_KEY": "your-api-key-here"
      }
    }
  }
}

Step 3: Restart Claude Desktop

Close and reopen the app. You will see "anonym-legal" listed under active servers.

Cursor IDE Setup

Cursor uses the same protocol. Add this to .cursor/mcp.json:

{
  "mcpServers": {
    "anonym-legal": {
      "url": "https://anonym.legal/mcp",
      "transport": "sse",
      "headers": {
        "Authorization": "Bearer your-api-key-here"
      }
    }
  }
}

What Gets Anonymized

The server catches 285+ entity types across 48 languages:

Example Transformation

Your prompt:

Debug this error from user john.smith@acme.com:
Payment failed for card 4532-1234-5678-9012
Customer ID: CUST-12345, IP: 192.168.1.100

What the model sees:

Debug this error from user [EMAIL_1]:
Payment failed for card [CREDIT_CARD_1]
Customer ID: [CUSTOMER_ID_1], IP: [IP_ADDRESS_1]

The reply you see:

The error for john.smith@acme.com suggests card
4532-1234-5678-9012 may have low funds...

You see real values. The model only ever saw tokens.

Advanced Options

Custom patterns — add your own regex to the CUSTOM_PATTERNS env var:

"CUSTOM_PATTERNS": "JIRA-[0-9]+,TICKET-[A-Z0-9]+"

Allowlist — keep public names from being masked:

"ALLOWLIST": "Anthropic,Claude,anonym.legal"

Disable entity types — let certain categories pass through:

"DISABLED_ENTITIES": "PHONE_NUMBER,URL"

Where Processing Happens

The proxy runs on your machine. Only the detection call goes to anonym.legal. Your prompts are not stored. See the privacy policy for details.

Pricing

The integration is included in all plans:

Most developers stay on Basic at €3/month.

Conclusion

AI tools are now part of daily dev work. They do not need to see your customers' data to be useful. The proxy handles that for you.

The integration:

• Needs no workflow changes
• Works with Claude Desktop, Cursor, and VS Code
• Guards PII on every prompt, every time
• Costs €3/month for most developers

Set it up once. Your data stays safe by default.

• Sign up for free
• MCP server docs
• View all entity types

When This Approach Has Limits

An MCP proxy makes secure AI usage close to frictionless, but it changes where your trust sits rather than removing the need for it — and a few cases fall outside its protection.

Detection happens server-side and is not perfect. The proxy sends your prompt to a detection service before forwarding clean text to the model. That detection step can miss an unusual identifier or an identifier embedded in code or logs, and anything it misses reaches the AI provider in the clear. For highly sensitive material, sanity-check the masked output rather than trusting every prompt blindly.

The proxy only protects traffic that flows through it. A teammate who pastes directly into the web UI, uses an AI surface you have not wired through MCP, or disables the server for a "quick question" bypasses the safeguard entirely. Protection is a function of consistent configuration across every tool and every developer, not of installing it once.

Tokens are reversible, so key handling matters. Reversible tokenization is what lets the model's reply come back with real values restored — but that means the mapping exists and is sensitive. Under the GDPR, tokenized output is pseudonymized, not anonymized; treat it as personal data and protect the key accordingly.

You still trust the detection service and your own machine. The proxy keeps raw text away from the model vendor, but the detection call goes to anonym.legal, and the proxy runs locally with your API key. A compromised developer machine or a leaked key undermines the model regardless of how well the proxy works. Pair it with normal key hygiene and endpoint security.

Sources

• Model Context Protocol specification
• LayerX — 77% of developers use AI tools with sensitive data