claude-code-slides

Hooks & Security — Setup Guide

Related slides: Slide 13 — Protect Your Secrets, Slide 14 — Control Agent Access

Official docs: Claude Code Hooks · Claude Code Security · Claude Code Permissions

Why Hooks?

Settings-based permissions are interpreted by the LLM — it can rationalize exceptions, especially in long sessions. Hooks are deterministic. They run shell commands at lifecycle points with binary pass/fail. No LLM involved, can’t be convinced otherwise.

PreToolUse   → runs BEFORE the agent acts     → exit 0 = allow, exit 2 = block
PostToolUse  → runs AFTER the agent acts      → validate output
Notification → runs on events                 → status line, alerts

Reference: paddo.dev/blog/claude-code-hooks-guardrails

Security Rules with Hookify

Hookify is an official Anthropic plugin that creates hooks from natural language — no manual JSON editing.

# Install hookify
claude plugin install --marketplace claude-plugins-official hookify

# Create a hook
> /hookify Block rm -rf commands targeting home directories

Pre-built Security Rules

These are hookify rules I run globally (~/.claude/). Each one is a .md file with a regex pattern and a block/warn action.

1. Block force push

name: block-force-push
event: bash
action: block
pattern: git\s+push\s+.*(-f|--force|--force-with-lease)

Force pushing rewrites remote history and can destroy teammates’ work. Do it manually if you truly need to.

2. Block rm -rf on home directories

name: block-rm-rf-home
event: bash
action: block
pattern: rm\s+-rf\s+.*(/Users/|~/|\$HOME|/home/)

Prevents catastrophic recursive deletion of home directory paths.

3. Detect hardcoded secrets in code

name: block-hardcoded-secrets
event: file
action: warn
conditions:
  - field: new_text
    operator: regex_match
    pattern: (API_KEY|SECRET|TOKEN|PASSWORD|PRIVATE_KEY|ACCESS_KEY|CLIENT_SECRET)\s*[:=]\s*['"][^'"]{8,}['"]

Warns when the agent writes what looks like a hardcoded credential. Suggests env vars or secrets managers instead.

4. Protect .env files (file access)

name: protect-env-files
event: all
action: block
conditions:
  - field: file_path
    operator: regex_match
    pattern: (^|/)\.env($|\.(?!example).*)

Blocks reading or editing .env, .env.local, .env.production, etc. Allows .env.example for documentation.

5. Protect .env files (bash commands)

name: protect-env-files-bash
event: bash
action: block
pattern: cat\s+.*\.env|less\s+.*\.env|head\s+.*\.env|tail\s+.*\.env

Catches bash commands that try to read .env files directly.

Creating Your Own Security Rules

# Natural language → hookify generates the rule
> /hookify Block any npm publish commands
> /hookify Warn when writing files outside the project directory
> /hookify Block database DROP TABLE commands

Or use /hookify without arguments to analyze your conversation history and suggest hooks for mistakes that happened.

Use Case: Status Line

Hooks can enhance your terminal UI. The Notification event fires on every status change, letting you render a custom status bar.

claude-hud

A visual HUD showing model info, context usage, active tools, and task progress.

claude-hud preview

ccstatusline

A powerline-style status formatter with widgets, emoji support, and auto-alignment.

ccstatusline demo

ccstatusline block timer

Because everyone deserves a nice terminal UI.

Use Case: Custom Search with Perplexity

Claude Code’s built-in WebSearch may not work behind corporate proxies. I created a hook-based plugin that intercepts search and routes it through Perplexity’s Sonar model — a more agentic search engine.

How it works: A PreToolUse hook intercepts WebSearch calls, sends the query to Perplexity Sonar via a LiteLLM proxy, and returns results — all before the default search runs.

# Add the marketplace
/plugin marketplace add https://github.tools.sap/I773117/claude-perplexity-search.git

# Install
/plugin install perplexity-search@claude-perplexity-search

# Configure
export PERPLEXITY_PROXY_URL="http://localhost:6655/litellm/v1/chat/completions"
export PERPLEXITY_API_KEY="your-key"

More Use Case Ideas

Hooks are a programmable extension point — the possibilities are endless:

Use Case Hook Event What It Does
Auto-format on save PostToolUse (Edit/Write) Run prettier/eslint after file edits
Lint gate PostToolUse (Edit) Block commits if linting fails
Cost tracking Notification Log token usage to a file or dashboard
Slack notifications PostToolUse Post to Slack when tasks complete
Auto-commit PostToolUse (Edit) Auto-commit after each successful edit
Screenshot on error PostToolUse (Bash) Capture terminal state when commands fail
Branch protection PreToolUse (Bash) Block operations on main/master branch
Token budget PreToolUse Warn when approaching context limit

The pattern is always the same: event → regex/code check → allow or block. If you can write a shell script for it, you can make it a hook.

Quick Reference

# Install hookify
claude plugin install --marketplace claude-plugins-official hookify

# Create hook from natural language
> /hookify Block dangerous rm commands

# Create hooks from conversation mistakes
> /hookify

# List all rules
> /hookify:list

# Enable/disable rules
> /hookify:configure

Further Reading