securitycheck

Block API keys, tokens, .env files, and private keys from leaking into git. Ships as a skill for Claude Code, OpenAI Codex CLI, Google Antigravity, and Moonshot Kimi CLI, plus a standalone CLI / git hook that works without any agent.

After the GitHub OAuth-token compromise in early 2026, "my repo is private" stopped being a credible secrets-management strategy. Any secret committed to a repo — public or private — should be considered compromised the moment it enters .git/objects. securitycheck runs before commits land so it never gets that far.

What it does

Three phases, always run all three:

1. .gitignore audit — verifies the file exists and covers .env, *.pem, *.key, SSH keys, .aws/, service-account JSON. Flags any sensitive file that's already tracked (a .gitignore rule does not retroactively untrack).
2. Staged-file check — refuses to commit a .env, private key, or credential file, regardless of content.
3. Diff content scan — applies ~30 provider-specific regexes to the staged diff. Catches AWS, GitHub, OpenAI, Anthropic, Google, Slack, Stripe, npm, DigitalOcean, HuggingFace, Azure, MongoDB/Postgres URIs, PEM private key blocks, JWT, and generic password=... assignments (under --strict).

Findings come back as BLOCK (exit 1, refuse commit) or WARN (looks like a placeholder or public sample — human verifies).

Install

As an agent skill

Recommended — via the Vercel Labs skills CLI (supports 55+ agent CLIs including Claude Code, Codex, Antigravity, Kimi, Cursor, OpenCode, Gemini CLI, Cline, Roo, Windsurf, Qwen Code, Goose, …; auto-detects which one you have):

# Install to whichever agent is detected on this machine
npx skills add SpectreMercury/404labs --skill securitycheck

# Install globally instead of project-local
npx skills add SpectreMercury/404labs --skill securitycheck -g

# Install to every supported agent on your system
npx skills add SpectreMercury/404labs --skill securitycheck --all

# Target specific agents
npx skills add SpectreMercury/404labs --skill securitycheck -a claude-code -a codex

Alternative — via this package's built-in installer (no extra CLI needed, but covers only the four listed below):

npx @404labs/securitycheck install                      # auto-detect, install to each
npx @404labs/securitycheck install --target all         # install for all 4 supported
npx @404labs/securitycheck install --target claude,kimi # comma-separated explicit list
npx @404labs/securitycheck install --list-targets       # show what's supported

All paths receive the same SKILL.md (YAML frontmatter + Markdown). Restart your CLI so the skill index picks it up.

For agents with no native skill loader — go through a host CLI:

• Zhipu GLM — distribute via GLM-skills / clawhub, or just call npx @404labs/securitycheck scan from your own pre-commit hook.
• MiniMax — use the MiniMax-AI/skills marketplace (it redistributes into Claude Code / Cursor) or wire the CLI in manually.

As a one-off CLI

npx @404labs/securitycheck scan          # scan the staged diff
npx @404labs/securitycheck scan --strict # also generic password=... heuristics
npx @404labs/securitycheck scan --all    # scan working tree, not just staged
npx @404labs/securitycheck scan --json   # machine-readable output

After a global install (npm i -g @404labs/securitycheck) the securitycheck binary is on your $PATH, so you can drop the npx @404labs/ prefix.

As a git pre-commit hook (no dependencies)

npx @404labs/securitycheck hook > .git/hooks/pre-commit
chmod +x .git/hooks/pre-commit

With Husky

npm install -D husky @404labs/securitycheck
npx husky init
echo 'npx securitycheck scan' > .husky/pre-commit

With lefthook

# lefthook.yml
pre-commit:
  commands:
    securitycheck:
      run: npx securitycheck scan

Output

securitycheck — pre-commit scan

  .gitignore:      present
  BLOCK findings:  2
  WARN findings:   1

  BLOCK — do not commit:
    • src/config.ts:14 — Anthropic API key [anthropic-key]
      const key = "sk-ant-api03-AbCdEf...";
    • .env — staged sensitive file
      Fix: git restore --staged ".env" && add to .gitignore

  WARN — verify these manually:
    • tests/fixtures/token.js:3 — JWT-shaped token [jwt]
      const t = "eyJhbGciOiJIUzI1NiJ9.eyJzdWIiOiIxMjM...";

  ✗ 2 blocking finding(s). Refusing to commit.
    Bypass (NOT recommended): SECURITYCHECK_SKIP=1 git commit ...

JSON mode (--json) emits the same data as a single JSON object — see test/scan.test.js for the shape.

Exit codes

What it catches

Full catalogue in skill/references/secret-patterns.md. Highlights:

Placeholder-like strings (YOUR_API_KEY, xxx, changeme, <API_KEY>) are auto-downgraded to WARN.

When a real secret is found

If the secret has ever been committed (not just staged), unstaging is not enough. In order:

1. Rotate at the provider immediately. Assume it's already compromised.
2. Remove from history — git filter-repo --path <file> --invert-paths or BFG. Squashing in a PR does not remove the blob.
3. Force-push. Coordinate with collaborators; this rewrites shared history.
4. Add the path to .gitignore so it can't come back.

Order matters. Cleaning a still-valid key buys nothing — the attacker already has it cached.

Configuration

No config file. Behaviour is controlled by CLI flags:

--all         Scan working tree, not just staged diff
--strict      Enable lower-confidence heuristics (more false positives)
--json        Machine-readable output
--no-color    Disable ANSI colors
--no-ignore   Skip the .gitignore audit
--no-files    Skip the sensitive-file presence check
--no-content  Skip the diff content scan

Environment:

• SECURITYCHECK_SKIP=1 — bypass entirely. Intended for emergencies only; the output makes the bypass visible in CI logs.

FAQ

Why not use gitleaks / trufflehog? Use them too if you can — they're battle-tested. securitycheck is zero-config, has no Go/Python dependency, and ships as a Claude Code skill so the agent inside your editor checks before you do. If a repo already runs gitleaks in CI, this is a strictly local belt-and-suspenders.

False positives? Three guards: provider-specific prefixes (we don't match sk- generically; we require sk-proj- / sk-ant- / 32+ chars and not a Stripe prefix); a placeholder heuristic that downgrades YOUR_API_KEY/xxx/changeme; and a --strict opt-in for the lossy generic password= rule.

False negatives? Yes, by design. We don't do entropy scanning in v0.1 — it has too many false positives without per-language tuning. Add provider-specific patterns via PR; see skill/references/secret-patterns.md.

Why ship a Claude Code skill at all? Because the agent writing your code is also the one most likely to paste a secret into it. Wiring securitycheck into the agent's pre-commit workflow closes that loop before the commit reaches your git index.

Roadmap

• Entropy-based detection (opt-in, per-file-type tuned)
• gitleaks-compatible config file consumption
• Pre-push hook variant that scans the full pushed range
• GitHub Action wrapper
• Per-project pattern overrides via .securitycheck.json

License

MIT — see LICENSE.