Coding agents

Coding agents commit on your behalf and often never run the Git hooks you installed. nosecrets hook claude-code plugs into the agent’s own hook system: before Claude Code runs a git commit, nosecrets scans what the commit is about to contain and blocks the call when it finds a secret. The reason goes back to the model, so it can fix the file and try again.

Setup

Add the hook to .claude/settings.json in the repository (or to ~/.claude/settings.json for all repositories). nosecrets must be on the PATH of the shell Claude Code runs in; see Installation.

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          { "type": "command", "command": "nosecrets hook claude-code" }
        ]
      }
    ]
  }
}

The hook reads the JSON that Claude Code sends on standard input (tool_name, tool_input.command and cwd). It does nothing for other tools and for commands that are not a git commit, and it does not scan in that case.

What is scanned

For a git commit, nosecrets runs the equivalent of nosecrets scan --staged in the repository of the command, with that repository’s .nosecrets.toml and .nosecretsignore. Which repository that is follows the command:

• git -C <dir> commit and cd <dir> && git commit use <dir>, relative to the cwd that Claude Code reports.
• git commit -a, -am and --all additionally scan the working-tree version of tracked files with unstaged changes, because the commit stages them.
• git add ... && git commit additionally scans the unstaged changes and the untracked files that are not ignored by Git, because the git add has not run yet when the hook is called. This covers more files than a narrow git add <path> would stage.

When it blocks

If a critical, high or medium finding is present, the hook prints each finding to standard error with path, line, rule ID, fingerprint and a masked preview, plus a hint on how to proceed, and exits with code 2. Claude Code then blocks the call and shows the text to the model. The secret itself is never printed.

To accept a finding on purpose, run nosecrets ignore <fingerprint> (optionally with --path) or put an @nsi marker on the line; see How it works.

Exit codes

The hook has its own contract, separate from the scan exit codes:

Code	Meaning
0	Not a commit, no blocking findings, or the hook could not check (a warning is printed to standard error).
2	Blocking findings; Claude Code blocks the command.

Errors inside the hook (invalid JSON on standard input, a directory that is not a Git repository, an invalid configuration, a Git or read error) print a warning and exit 0. Claude Code treats every other non-zero code as a non-blocking error anyway, and a misconfigured hook should not stop the agent from working. The cost is that a commit goes through unchecked when the hook cannot run, so keep nosecrets scan --staged as a pre-commit hook for your own commits.

Limits

• The command is split into words on &&, ||, ;, |, & and line breaks. Text in quotes is one word, so echo "git commit" is not a commit. This is a lexical check, not a shell parser.
• Commits that are not spelled out in the command are not seen: bash -c 'git commit', $(...), scripts, aliases, shell functions and git commit run by another tool.
• cd with ~, variables or - is not resolved. git commit <path> (commit only those files) is checked like a plain git commit.
• Only the Bash tool is checked, not file writes or other tools.
• The settings snippet is documented from the Claude Code hooks reference.