ClaudeMods
☰
JA
● 0 人がオンライン ・閲覧 0 回
スポンサー作品を投稿
GitHub リポジトリ · 投稿者 0xGondarxyz

secret-guard

秘密情報を会話と公開 repository から遠ざける Claude Code mod。

0xGondarxyz@0xGondarxyz

0xGondarxyz/claude-code-mods/tree/main/secret-guard

翻訳済み

この mod について

secret-guard は Claude が読む前に、すべてのツール結果の API key、token、パスワードをマスクし、秘密情報やホームパスを公開 repository に漏らす git push を阻止します。接頭辞付き token(Anthropic、OpenAI、GitHub、AWS、Google、Slack、Stripe、GitLab、npm、Telegram、JWT など)、PEM 秘密鍵ブロック、Bearer token、URL パスワード、秘密らしい代入を検出し、それぞれを種類と末尾 4 文字を示すマーカーに置き換えます。書き込みガードは、マーカーを含む Write/Edit/MultiEdit/NotebookEdit と Bash コマンドを拒否します。push チェックは repository と remote を解決し、gh に公開範囲を問い合わせ、未 push の commit の追加行(コマンドが add または commit も行う場合は作業ツリーの diff と未追跡ファイルも)をスキャンします。秘密情報または絶対ホームパスが見つかると、最大 10 件の file:line 指摘を付けて push を拒否します。/plugin marketplace add 0xGondarxyz/claude-code-mods と /plugin install secret-guard@claude-code-mods でインストールするか、--plugin-dir で実行します。MIT ライセンス。mod はサンドボックス化されません。

インストール

まず作者の README で marketplace とプラグイン名を確認してください。コマンドはリポジトリの構成によって変わる場合があります。

claude plugin marketplace add 0xGondarxyz/claude-code-mods
claude plugin install secret-guard
原文 / README

secret-guard

A Claude Code mod that keeps secrets out of the conversation and out of public repos.

  1. It masks API keys, tokens and passwords in every tool result before Claude reads them.
  2. It blocks a git push that would leak a secret or a home path to a public repo.

Install

/plugin marketplace add 0xGondarxyz/claude-code-mods
/plugin install secret-guard@claude-code-mods

Or try it without installing:

git clone https://github.com/0xGondarxyz/claude-code-mods
claude --plugin-dir claude-code-mods/secret-guard

Mods are not sandboxed. They run with the same access as Claude Code. Read the source before you install any mod, including this one.

Masking

Every tool result is checked before it is stored: built-in tools, MCP tools, the main thread and subagents. Each secret becomes a marker. The rest of the text is kept byte for byte.

ANTHROPIC_API_KEY=[masked anthropic_key ...a1b2]

The marker shows the kind and the last 4 characters.

What it finds:

  • Prefixed tokens: Anthropic, OpenAI, GitHub (ghp_, gho_, ghu_, ghs_, ghr_, github_pat_), AWS AKIA, Google AIza, Slack, Stripe, GitLab, Apify, Notion, Hugging Face, Replicate, npm, Telegram bot tokens, JWTs.
  • PEM private key blocks (the whole block).
  • Bearer tokens (Bearer followed by 20 or more token characters).
  • The password in a URL such as postgres://user:PASSWORD@host.
  • Assignments: a name that contains api_key, secret, token, password, credential, private_key, access_key or auth, then = or :, then a value of 16 or more characters with a letter and a digit. Only the value is masked.

What it leaves alone: placeholders (your_..., xxx..., <...>, ${...}, process.env..., os.environ..., changeme, example), git SHAs, UUIDs, numbers, file paths, npm sha512- integrity hashes.

Claude now sees markers, so it could write one back over the real secret. The write guard stops that. Write, Edit, MultiEdit and NotebookEdit are denied when the new text holds a marker, and so is a Bash command that holds one. Claude is told to edit around the line or ask you to change that value.

Push check

When a Bash command runs git push, secret-guard checks before it runs:

  1. It finds the repo (git -C <dir>, a leading cd <dir> &&, or the session folder) and the remote (default origin).
  2. It asks gh repo view for the visibility. PRIVATE or INTERNAL: the push goes through, no scan. PUBLIC or unknown (no gh, not GitHub, an error): it scans.
  3. It scans the added lines of every local commit the remote does not have. If the same command also runs git add or git commit, it scans the working tree diff and the untracked files too (not ignored, under 1 MB, not binary).
  4. It looks for the same secrets as masking, plus absolute home paths: your real home folder and any /home/<name>/ or /Users/<name>/ path, and Windows paths (C:\Users\<name>\, the forward-slash form and the JSON-escaped form, any drive letter). HOME and, when set, USERPROFILE count as your home folder. /home/user/, /home/runner/ and the Windows names Public, Default, Default User and All Users are ignored.

With findings, the push is denied. The text lists up to 10 as file:line kind. It never prints the secret. A toast says secret-guard: push blocked, N findings. When the visibility was unknown, the text says so.

If the scan itself fails (git error, timeout of about 5 seconds), the push is allowed and a toast says secret-guard: scan failed, push allowed.

Commands

| Command | What it does | | --- | --- | | /secret-guard | Shows how many values were masked this session and whether a pass is active. | | /secret-guard allow | Lets the next push skip the scan. The pass lasts 10 minutes and works once. |

For a false positive: run /secret-guard allow, then push again.

Limits

  • Detection is by pattern. A secret with no known prefix and no telling name (for example a bare random string) is not masked.
  • What the model reads is masked. The engine stores a tool result's structured record (what the screen draws) as made, so the transcript file on disk can still hold the raw output.
  • Rows that are not tool results (your prompts, attachments) are not masked.
  • git log output is cut at 4 MB. A very large unpushed history is only scanned in part.
  • Pushes started outside Claude Code's Bash tool (hooks, scripts it launches) are not seen.
  • A file that must contain the literal marker text cannot be written by Claude. Ask Claude to leave it to you.

No options.

License

MIT

同名の他の作品

関連作品