@claude-code-hooks/secrets

A Claude Code hook package that helps users feel safer by warning when secret-like tokens appear in tool inputs — and optionally scanning staged git files before commits.

Language / 언어: English (main) · 한국어 (sub) — 한국어 안내는 루트 문서 하단에 포함되어 있습니다.

This is intentionally high-signal and lightweight: it detects only a few patterns that are commonly accidental leaks.

Part of the claude-code-hooks monorepo.

Contributing · Security Policy · Releasing · License

Install / run

npx @claude-code-hooks/secrets@latest

Project config: claude-code-hooks.config.json

Uses the same project config file as other packages.

Minimal schema:

{
  "secrets": {
    "mode": "warn",
    "enabledEvents": ["PreToolUse", "PermissionRequest"],
    "scanGitCommit": false,
    "ignore": { "regex": [] },
    "allow": { "regex": [] }
  }
}

• scanGitCommit: when true, intercepts git commit commands on PreToolUse and scans staged files (via git diff --cached) for secret patterns. Default: false.
• allow.regex: if any pattern matches, findings are suppressed
• ignore.regex: if any pattern matches, findings are suppressed and a dim note is printed

Modes

• warn (default): prints warnings to stderr, exits 0
• block: exits 2 only for HIGH confidence findings (private key material)

Git commit scanning

When scanGitCommit is enabled and Claude tries to run a git commit command (detected via PreToolUse), the hook will:

1. Run git diff --cached --name-only --diff-filter=ACM to list staged files
2. Skip binary files and excluded directories (node_modules, .git, dist, etc.)
3. Scan each file against an expanded set of secret patterns
4. Merge file-level findings with the standard payload scan

File-level patterns include everything from the payload scanner plus:

• Google API keys (AIza...)
• Stripe secret keys (sk_live_..., rk_live_...)
• Twilio API keys
• SendGrid API keys (SG.xxx.xxx)
• npm tokens (npm_...)
• PyPI tokens (pypi-...)
• Database connection strings with embedded credentials
• Generic secret=, token=, password= assignments with high-entropy values

Blocking behaviour is the same: only HIGH severity findings (private key material) trigger exit 2 in block mode.

Commands

claude-secrets list-events
claude-secrets run --event PreToolUse --mode warn
claude-secrets doctor

What it detects (v0.1)

• HIGH: -----BEGIN (RSA|OPENSSH|EC|PGP|DSA|ENCRYPTED) PRIVATE KEY-----
• MED: OpenAI sk-..., GitHub ghp_... / github_pat_... / gho_... / ghu_... / ghs_... / ghr_..., AWS AKIA..., Slack xox*, Google AIza..., Stripe sk_live_..., Twilio SK..., SendGrid SG...., npm npm_..., PyPI pypi-..., DB URLs with credentials, generic secret assignments

Sanity checks

# warn
echo '{"text":"sk-1234567890123456789012345"}' | npx --yes @claude-code-hooks/secrets@latest run --event PreToolUse

# block (HIGH only)
echo '{"text":"-----BEGIN OPENSSH PRIVATE KEY-----"}' | npx --yes @claude-code-hooks/secrets@latest run --event PreToolUse --mode block

# git commit scan (requires scanGitCommit: true in config)
echo '{"tool_input":{"command":"git commit -m test"}}' | npx --yes @claude-code-hooks/secrets@latest run --event PreToolUse