claude-code-privacy-guard

Claude Code Privacy Guard

🛡️ Prevent secrets and PII from being accidentally shared with Claude Code.

A privacy-first plugin for Claude Code that scans prompts for sensitive data and blocks them before they reach the AI.

Features

• ✅ Blocks prompts containing sensitive data before they're sent to Claude
• ✅ Detects PII, secrets, API keys, tokens, and sensitive information
• ✅ Works locally - all scanning happens on your machine
• ✅ Zero configuration - works out of the box
• ✅ Detailed reporting - shows exactly what was detected
• ✅ Toggle rules and manage allowlists - by hand in .privacy-guard.json, or via npx claude-code-privacy-guard rules for a local web UI

Installation

# Add the marketplace (if not already added)
/plugin marketplace add datumbrain/claude-code-privacy-guard

# Install the plugin
/plugin install claude-code-privacy-guard

⚠️ Important: Restart Required

After installing the plugin, you must restart your Claude Code session for it to take effect. This is because hooks are registered at session startup - Claude Code doesn't dynamically load new hooks mid-session.

Simply close and reopen Claude Code, or start a new session.

Once restarted, the plugin will automatically scan all prompts before they reach Claude.

What Gets Detected

✅ Secrets

• OpenAI API keys (sk-..., sk-proj-...)
• Anthropic, OpenRouter, Google AI, Groq, and Perplexity API keys
• Hugging Face API tokens
• AWS credentials
• GCP service account keys
• Azure client secrets (Microsoft Entra ID)
• GitHub tokens (classic and fine-grained PATs)
• GitLab tokens (personal access, deploy, runner, pipeline trigger, service account, CI job)
• Slack tokens
• Stripe keys
• Twilio API key SIDs and account SIDs
• SendGrid API keys
• npm access tokens
• Database connection strings with inline credentials (postgres/mysql/mongodb/redis/amqp)
• JWT tokens
• Bearer tokens
• SSH and PEM private keys
• Generic API key/secret/token assignments in code

Plus a bundled external ruleset (data/regex_list_1.json, filterable via externalRulesMode) covering many more service-specific tokens - see Managing Rules to browse the full list.

✅ Personal Information (PII)

• Email addresses

Phone numbers, Social Security Numbers, and credit card numbers are not currently detected by the default configuration - see #7 for tracking status.

How It Works

1. You type a prompt in Claude Code
2. Privacy Guard intercepts it via a UserPromptSubmit hook
3. Scans for sensitive data using regex patterns
4. Reacts according to the configured mode (see Block, Redact, or Warn? below) - by default, blocks the prompt if sensitive data is found
5. Shows you exactly what was detected

Blocking relies on the UserPromptSubmit hook JSON protocol: the hook prints {"decision": "block", "reason": "..."} to stdout and exits with code 0. (Exit code 0 is required for the JSON decision to be honored - a non-zero exit is treated as a non-blocking hook error, and the prompt would go through anyway.) A non-blocking warning uses {"systemMessage": "..."} instead, which surfaces the message without stopping the prompt.

Example

Input:

My API key is sk-proj-abc123xyz and email is [email protected]

Result:

🛡️ Privacy Guard blocked this prompt

Found 2 sensitive item(s):
  - OpenAI API Key (openai-api-key): sk-p…3xyz
  - Email Address (email-address): john….com

Risk Score: 100/100
Secrets: 1 | PII: 1

Please remove or anonymize sensitive data before proceeding.
To disable a rule, add its ID to "disabledRules" in .privacy-guard.json.

Configuration

Create a .privacy-guard.json file (searched upward from your current working directory, so a repo-root or home-directory config both work) to override the defaults. See .privacy-guard.example.json for a starter file.