Scan Secrets

ggshield — GitGuardian Secret Scanner

Overview

ggshield is a CLI that detects 700+ types of hardcoded secrets — AWS keys, GitHub tokens, database connection strings, private keys, Stripe keys, Slack webhooks, JWTs, and more — in files, git history, Docker images, and PyPI packages.

Core rule: when working on code that handles credentials, run ggshield before presenting the result. Do not commit or surface code that contains a detected secret.

Start Here — Read This Before Doing Anything

Do not skip this section.

• Do not improvise alternate scanners. No grep one-liners, no regex hunts, no custom secret-finding scripts. Use ggshield secret scan with the flags documented in Scan commands below. The detectors are tuned and validated; ad-hoc patterns are not.

• The ggshield CLI is mandatory for scanning — do not use the GitGuardian Developer MCP scan_secrets tool as a substitute. If the MCP server is connected, its scan_secrets tool will be tempting as a no-install shortcut. It is the wrong tool for this skill: it scans a single in-memory payload you paste in, so it is slow for anything larger than a snippet and cannot scan git history, commit ranges, staged changes, repositories, Docker images, or PyPI packages — which is the core of what this skill does. The CLI streams files locally and audits full history in one pass; the MCP path cannot. So:

  • Never silently fall back to scan_secrets (MCP) because the CLI isn't installed yet.
  • If ggshield is not installed, strongly recommend the user install it before scanning — one command, under a minute, and it unlocks history/commit/Docker/PyPI scanning the MCP tool can't do. Run Onboarding (below) and make the case rather than reaching for the MCP shortcut. Only if the user explicitly declines to install should you note that MCP scan_secrets exists, and even then only for a single pasted snippet — never for history, a repo, or any command in Scan commands.

• Do not improvise remediation advice. No general-knowledge rotation walkthroughs, no improvised git filter-repo / BFG suggestions, no HMSL omissions. When ggshield returns one or more findings, read references/remediation-doctrine.md before composing any user-facing remediation message — the doctrine differs from common defaults in important ways (rotation > history rewrite; HMSL is the prescribed follow-up for unverifiable validity).

• Always pass --json in agent contexts — you need structured output to parse findings reliably.

• Always pair -r with -y — -r triggers an interactive Confirm recursive scan. prompt that hangs on stdin without -y.

• Run Onboarding first if the CLI isn't set up. If ggshield --version fails or ggshield api-status errors, follow references/ggshield-cli-setup.md before attempting any scan, and strongly recommend the user install it — do not reach for the MCP scan_secrets tool as a workaround (see the CLI-is-mandatory rule above). Every scan command is useless until the CLI is installed and authenticated.

• Do not surface code containing a detected secret. Let the scan finish first. Do not begin remediation on the first hit — ggshield reports the complete finding set in one run, and the same credential often appears across several files or commits. Only once the scan has completed:

  1. Stop. Enumerate every finding, then group them: collapse the same credential value seen across multiple files / commits / artifacts into a single item, and keep distinct credentials separate. Report the grouped set (file(s), line(s), secret type, validity).
  2. Read references/remediation-doctrine.md end-to-end — do not skip this step. Common defaults on history rewriting, rotation triggers, and HMSL follow-up diverge from GitGuardian doctrine.
  3. Triage the complete, deduplicated set, then compose one consolidated remediation plan: rotation first, HMSL follow-up for unverifiable-validity findings, history-rewrite only under the narrow conditions listed in the reference. One credential is one rotation, even if it leaked in five places. A valid validity result does not let you skip triage — it raises urgency, but ownership and blast radius still select the remediation shape, and a valid production-critical credential is a supervised, sequenced rotation (Coordination), not a fast "just rotate it." See doctrine principle 7.

  Do not commit, do not show the code with the secret inline, do not "just continue and we'll fix it later," and do not start rotating one finding while others are still being scanned or triaged.

• Do not extend this skill's agent-executable contract to HMSL. When a finding's validity is unknown, cannot_check, or no_checker, the natural follow-up is HasMySecretLeaked (HMSL) — GitGuardian's privacy-preserving hash-lookup service for known credentials against the public-leak corpus. HMSL has a different execution model — user-run only — and the contract holds whether or not the user has the dedicated check-hmsl skill installed:

  • Do not invoke ggshield hmsl check, fingerprint, query, decrypt, or check-secret-manager yourself.
  • Do not read the credential file with Read / Grep / cat / head / tail / sed / awk / less / xxd / wc / file / ls or any other tool — that pulls plaintext into the agent context before HMSL's local-hashing protocol can protect it.
  • Print the exact command for the user to run in their own terminal. Use -n none --json so the output the user pastes back contains no identifying hints. Suggest ggshield hmsl quota before any bulk check. Refuse --naming-strategy cleartext outright.
  • If the user has the check-hmsl skill installed, the agent should load it for the full protocol and command set; if not, the rules above are sufficient on their own — do not gate the follow-up on having the other skill.

When to Use

Trigger a scan when:

• The user asks to scan a file, directory, or repository for secrets or credentials
• You are writing or modifying code that handles API keys, tokens, passwords, connection strings, or any credentials — scan before presenting the result
• The user is about to commit or push — scan staged changes first

What ggshield covers:

• Scan files and directories on disk (secret scan path)
• Audit a repository's full git history (secret scan repo) — catches secrets committed and later deleted
• Scan a specific commit, a commit range, or staged changes
• Scan Docker images and PyPI packages
• Run as a CI gate that fails on findings
• Install git hooks (pre-commit / pre-push) and AI agent hooks (Claude Code, Cursor, Copilot) — the agent hooks scan the prompt, tool calls, and tool outputs from inside the agent itself
• Manage false positives via # ggignore comments and .gitguardian.yaml

For detailed command variants, expected JSON output shapes, and CI integration, see references/workflows.md. For interpreting scan output, the HMSL follow-up contract, and false-positive workflows, see references/interpreting-results.md. For remediation — triage, rotation rules, when (and when not) to rewrite git history, per-secret-type runbooks, and validation, see references/remediation-doctrine.md. For shared ggshield install, authentication, headless setup, CI tokens, and hook-install commands, see references/ggshield-cli-setup.md. For platform-wide topics that span every GitGuardian skill (public docs URL pattern, auth/scope recovery, instance URLs, headless setup), see references/gitguardian-platform.md.

When Not to Use

Do not use this skill when:

• The user already holds a known credential and wants to know whether it has leaked publicly — use check-hmsl. This skill finds unknown secrets; that one checks known ones against the public-leak corpus.
• The request is to inventory credentials across the whole machine (dotfiles, cloud CLI configs, shell history, other repos) — use scan-machine.
• You are tempted to substitute a grep one-liner, regex hunt, or custom secret-finding script. Use ggshield secret scan — the detectors are tuned and validated; ad-hoc patterns are not.

Onboarding (first use)

Prerequisites

• ggshield 1.49.0 or later — required for full feature support including the AI agent hooks (ggshield install -t claude-code, -t cursor, -t copilot). Older ggshield versions can scan but can't install agent hooks.
• A GitGuardian account (free tier available at https://dashboard.gitguardian.com/signup).

Setup

If ggshield --version succeeds and ggshield api-status returns OK, skip shared setup. Otherwise follow references/ggshield-cli-setup.md and return here once both checks pass.

Step 1 (recommended) — Install the agent hook for defense in depth

Once ggshield is installed and authenticated, the recommended first action on a new machine is installing the agent hook. The hook scans prompts, tool inputs, and tool outputs from inside the agent for detected secrets and blocks them before they reach the model context — defense in depth against the agent inadvertently reading or echoing a credential. Match the user's agent:

ggshield install -t claude-code -m global     # Claude Code
ggshield install -t cursor -m global          # Cursor
ggshield install -t copilot -m global         # Copilot

Propose this on first use with a one-liner: "For defense in depth — so credentials in files I read never reach my transcript — install the agent hook now? (ggshield install -t claude-code -m global)" Wait for the user's yes/no. The hook is one layer of defense; the user-run-only rule for ggshield hmsl * documented in Start Here is the other. The two are complementary, not substitutes.

Step 2 — Brief the user on what this skill enables

• Scan code for hardcoded secrets — automatically when handling credentials, or on request for a specific file or directory.
• Audit a repository's git history, a commit range, a single commit, a Docker image, or a PyPI package for leaked secrets.
• Block secrets before they are written via the agent hook installed in Step 1; or via git hooks (ggshield install --mode local) and ad-hoc scans of staged changes (scan pre-commit).
• Manage false positives via inline # ggignore comments or .gitguardian.yaml rules.

Keep the brief tight; the detailed setup reference is for the agent to consult, not for the user to read.

Scan commands

Always pass --json for structured output. Recursive scans (-r) trigger an interactive Confirm recursive scan. prompt — pair -r with -y whenever an agent invokes it, otherwise the CLI hangs waiting on stdin.

ggshield secret scan repo . --json                       # full git history
ggshield secret scan path -r -y . --json                 # current files, no git required
ggshield secret scan path <file> --json                  # single file (no -r needed)
ggshield secret scan pre-commit --json                   # staged changes
ggshield secret scan commit-range HEAD~5..HEAD --json    # commit range
ggshield secret scan commit <sha> --json                 # specific commit
ggshield secret scan docker <image> --json               # Docker image
ggshield secret scan pypi <package> --json               # PyPI package

Key flags:

Exit codes: 0 = no secrets found, 1 = secrets detected, 128 = unexpected error.

When findings are present — quick reference

Full doctrine in references/remediation-doctrine.md — load it before composing any user-facing remediation message. Dispatch by which command produced the finding:

Custom remediation message takes the lead. If the workspace has a configured remediation workflow, ggshield (≥ 1.30.0) prints the workspace's own message in its hook output. Surface that message verbatim as the primary guidance and fill in the doctrine mechanics around it — do not override it with generic advice. See § 13.

The three triggers most often missed:

• Pushed to a remote → rotate. History rewriting is generally discouraged once secrets are pushed; rotation is the actual remediation. Do not lead with git filter-repo / BFG.
• Validity is unknown / cannot_check / no_checker / failed_to_check → propose HMSL as the natural follow-up to check public leakage. Prepare the command (ggshield hmsl quota, then ggshield hmsl check ... -n none --json); the user runs it. Never invoke ggshield hmsl * yourself, never read the credential file.
• Local, never pushed → remove, don't rotate. Rewriting unpushed history is cheap and worth doing here.

Best Practices

• Scan proactively when writing or modifying code that handles credentials or configuration — do not wait to be asked.
• When a credential is found: always remove it from the code. Rotation is only necessary if the secret has been exposed on a remote — pushed to a shared repository, CI system, or any external service. A secret that is purely local and has never left the machine does not need rotation, only removal.
• Do not commit or present code that contains a detected secret. Stop the workflow, report the finding (file, line, secret type, validity), then fix and re-scan.
• For false positives, add # ggignore on the offending line, or run ggshield secret ignore --last-found to record it in .gitguardian.yaml.

Troubleshooting

ggshield: command not found — ggshield is not on PATH. See Onboarding (first use) above.

401 Unauthorized — the API key or stored OAuth token is missing or invalid. Verify with ggshield api-status. If using GITGUARDIAN_API_KEY, confirm the value with echo $GITGUARDIAN_API_KEY and that the token has the scan scope.

403 Forbidden / "Insufficient permissions" — the token is valid but is missing a scope this action requires. See references/gitguardian-platform.md for the recovery flow — ggshield auth logout + ggshield auth login --scopes <scope>, runnable on the user's behalf, no manual PAT creation needed.

Not a git repository — ggshield secret scan repo requires a git context. Use ggshield secret scan path -r -y . instead.

Recursive scan hangs — -r was used without -y. The CLI is waiting on the Confirm recursive scan. prompt. Re-run with -y.

OAuth browser window does not open — the environment is headless. Lead with ggshield auth login --method oob (ggshield 1.51.0+): it prints a URL the user opens on any device, signs in, and pastes the code back — no manually-created token. Fall back to --method token only if oob is unsupported (older ggshield, or an instance that doesn't offer it). See Headless and CI in references/ggshield-cli-setup.md.

Rate limiting — free tier quota exceeded. Direct the user to check usage at https://dashboard.gitguardian.com.

Any other or unlisted error — before improvising a fix, consult GitGuardian's AI-agent docs index at https://docs.gitguardian.com/llms.txt to locate the relevant page, then append .md to that page's URL to read it as Markdown. Search there first rather than guessing.