secret-tripwire

PostToolUse hook for Claude Code that scans tool outputs for leaked secrets before they enter the model's context window. Once a secret hits the context, it's effectively leaked to the model. This intercepts it first.

Install

npm install -g secret-tripwire

Setup

Add to ~/.claude/settings.json:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": ".*",
        "hooks": [
          {
            "type": "command",
            "command": "secret-tripwire",
            "timeout": 5
          }
        ]
      }
    ]
  }
}

What It Detects

| Secret Type | Pattern | Example | |-------------|---------|---------| | AWS Access Key | AKIA* / ASIA* | AKIAIOSFODNN7EXAMPLE | | GitHub Token | ghp_, gho_, ghs_, ghr_, gha_, github_pat_ | ghp_ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefgh | | Private Key | -----BEGIN * PRIVATE KEY----- | RSA, EC, PKCS#8, PGP, OPENSSH, DSA | | JWT Token | eyJ...eyJ... (base64url, incl. alg:none) | eyJhbGciOiJIUzI1NiJ9.eyJzdWIiOi... | | Connection String | postgresql://, postgres://, mongodb+srv://, etc. | postgres://user:pass@host/db | | API Key | api_key: "...", secret_key=... | api_key: "sk_live_abcdef1234567890" | | Bearer Token | Bearer <40+ chars> | Bearer eyJhbGciOiJSUzI1NiIsInR5cCI6... |

How It Works

1. Claude Code calls a tool (Read, Bash, WebFetch, etc.)
2. The tool returns output
3. Before the output enters Claude's context, this hook scans it
4. If secrets are detected, a warning is injected via additionalContext
5. Claude sees the warning and knows not to repeat the raw output

This is a PostToolUse hook — it's advisory (exit 0 always). It cannot block tool execution, only warn after the fact. The warning tells Claude that secrets were detected and the raw output should not be referenced.

Library Usage

import { scan, redact } from 'secret-tripwire'

// Scan text for secrets
const result = scan('My AWS key is AKIAIOSFODNN7EXAMPLE')
console.log(result.clean)       // false
console.log(result.detections)  // [{ type: 'AWS_KEY', match: 'AKIAIOSF...', index: 14 }]

// Redact secrets from text
const { redacted, count } = redact('Key: AKIAIOSFODNN7EXAMPLE')
console.log(redacted) // 'Key: [REDACTED-AWS_KEY]'
console.log(count)    // 1

Design

• Zero false positives over zero false negatives — patterns are conservative (high-confidence only)
• Truncated matches — detected secrets are truncated to 8 chars in logs so the warning itself doesn't leak
• Always exits 0 — PostToolUse hooks are advisory, never blocking
• Zero runtime dependencies beyond @valencets/resultkit

Requirements

• Node.js >= 22
• ESM only

License

MIT