Claude Code Toolkit

Hooks and daemons that make working with Claude Code safer and lower-friction: secrets get scrubbed off your disk before they leak, per-directory READMEs stay fresh for agents, and identity guards keep the wrong GitHub account off the wrong repo.

Each tool stands alone — install only what you want. Clone once; the installers wire hooks into your repos and a scheduled scrubber onto your machine.

---

What's in here

tool	what it does	trigger
secret-scrub	Detects + redacts secrets — Solana keys, BIP39 mnemonics, GitHub/OpenAI tokens, PEM blocks, *_SECRET= env vars — across four leak pathways	git pre-commit + scheduled scans
docgen	Generates a one-paragraph README per directory that Claude auto-loads as context when it reads files there	git pre-push + Claude Code PreToolUse
identity-guard	Blocks pushes/merges/PRs/releases under the wrong GitHub account (personal identity on a bot repo, or vice-versa)	Claude Code PreToolUse

The four secret leak pathways secret-scrub covers

Most secret scanners catch one. A secret you handle while working with an AI agent can leak from any of these — so secret-scrub watches all four:

1. Staged files at commit time → pre-commit hook
2. Claude session transcripts (~/.claude/projects/*.jsonl) → scheduled scan
3. Working-tree files — uncommitted/untracked secrets the pre-commit hook never sees → scheduled scan
4. .git/config — tokenized remote URLs (https://user:TOKEN@host) that git remote -v echoes into logs → scheduled audit

Pathway 4 is the one almost everything misses, because .git/config lives inside .git/ where git status can't see it.

---

Quick start — let Claude set it up

Open Claude Code and paste:

Help me set up the Claude Code Toolkit from
https://github.com/decider/claude-code-toolkit — I'm not a developer,
so keep it plain-English, audit it for safety first, ask before each
install, and then run it for me.

Claude will clone it, walk the safety audit, and install the pieces you choose. Expected time: ~5 minutes.

Quick start — do it yourself

git clone https://github.com/decider/claude-code-toolkit
cd claude-code-toolkit

# Read the audit first — this code handles your secrets
less docs/SAFETY.md

# Interactive installer — asks before each component, installs nothing silently
./install.sh

Or install one tool at a time:

./secret-scrub/install.sh          # pre-commit hook (point at any repo)
./secret-scrub/install-launchd.sh  # macOS: scan ~/.claude/projects every 30 min
                                   # (Linux: install-systemd.sh)
./docgen/install-push-hook.sh      # auto-refresh per-dir READMEs on push
./identity-guard/install.sh        # wrong-account guard

---

Safety

This toolkit handles your secrets, so it's built to be verified, not trusted. docs/SAFETY.md is a checklist you (or Claude) run before installing anything:

• No outbound network in the scrubber — it's local-only
• No eval/exec on the file contents it scans
• Writes only to declared paths (.git/hooks/, scheduler dirs, the files being redacted)
• Install scripts do exactly what they advertise — no hidden hooks

A clean clone passes all of them. If a fork added something the checklist would catch, that's the signal to stop.

---

What it will NOT do for you

• Rotate an already-leaked secret. If the scrubber finds a live key, it redacts the on-disk copy — but that key is compromised and you must rotate it yourself.
• Catch secrets pasted into the browser app. It covers the Claude Code CLI's on-disk state, not server-side chat history.

---

Requirements

• python3 (3.9+) — secret-scrub
• node (18+) — docgen
• git, gh (GitHub CLI) — identity-guard, docgen push hook
• macOS (launchd) or Linux (systemd --user) — scheduled scrubbers

License

MIT — see LICENSE.