Hooks: Concrete Patterns

Auto-format after editing

PostToolUse hook on Edit|Write. Detects the extension of the modified file and applies the appropriate formatter.

INPUT=$(cat)
FILE=$(echo "$INPUT" | jq -r '.tool_input.file_path')
EXT="${FILE##*.}"
case "$EXT" in
  ts|tsx) npx prettier --write "$FILE" 2>/dev/null ;;
  py)     black "$FILE" 2>/dev/null ;;
  go)     gofmt -w "$FILE" 2>/dev/null ;;
  rs)     rustfmt "$FILE" 2>/dev/null ;;
esac
exit 0

Configure as async: true — formatting is cosmetic and should not block Claude.

macOS notification on completion

Stop hook. Triggers a macOS system alert with a contextual sound (success, error, waiting).

INPUT=$(cat)
MESSAGE=$(echo "$INPUT" | jq -r '.message // "Done"')
osascript -e "display notification \"$MESSAGE\" \
  with title \"Claude Code\""
afplay /System/Library/Sounds/Hero.aiff &
exit 0

Security: block dangerous commands

PreToolUse hook on Bash. Intercepts destructive patterns before execution.

COMMAND=$(echo "$TOOL_INPUT" | jq -r '.command')
DANGEROUS=("rm -rf /" "dd if=" "DROP DATABASE"
           "git push --force.*main")
for pattern in "${DANGEROUS[@]}"; do
  if [[ "$COMMAND" == *"$pattern"* ]]; then
    echo "BLOCKED: $pattern" >&2
    exit 2
  fi
done
exit 0

Exit code 2 blocks the action and surfaces the stderr message to Claude.

Git auto-stage after editing

PostToolUse hook on Edit|Write. Automatically prepares modified files for the next commit.

INPUT=$(cat)
FILE=$(echo "$INPUT" | jq -r '.tool_input.file_path // empty')
[[ -z "$FILE" || "$FILE" == "null" ]] && exit 0
git add "$FILE" 2>/dev/null
exit 0

RTK auto-wrapper

PreToolUse hook on Bash. Detects commands optimizable by RTK and informs Claude of the available alternative.

The logic: if the command starts with git log, cargo test, pnpm list or verbose equivalents, suggest rtk <command> via stdout before letting it through.

Log all executed commands

PostToolUse hook on Bash. Useful for auditing and debugging long sessions.

COMMAND=$(cat | jq -r '.tool_input.command // empty')
echo "$(date -u +%FT%T) | $COMMAND" \
  >> ~/.claude/logs/commands.log
exit 0

Dispatcher pattern: a single entry point

Rather than multiplying entries in settings.json, a dispatch.sh script routes to specialized handlers based on the file or tool. Result: a single Edit|Write|Bash matcher in the config, with separately maintainable handlers in .claude/hooks/handlers/.

Output replacement: redact secrets before Claude sees them (v2.1.121)

PostToolUse now supports output replacement for ALL tools. If your hook prints a JSON with an "output" key, Claude receives that instead of the real tool result. Use this to strip secrets from file reads.

INPUT=$(cat)
TOOL=$(echo "$INPUT" | jq -r '.tool_name')
OUTPUT=$(echo "$INPUT" | jq -r '.tool_response.content // empty')

# Redact .env values before Claude sees them
if [[ "$TOOL" == "Read" ]]; then
  CLEAN=$(echo "$OUTPUT" | sed 's/\(API_KEY=\).*/\1[REDACTED]/')
  echo "{\"output\": $(echo "$CLEAN" | jq -Rs .)}"
fi
exit 0

Invoking an MCP tool from a hook (v2.1.118)

Hooks can use type: "mcp_tool" to call any connected MCP server without a shell script. Example: post a Slack notification via an MCP Slack server whenever a file is edited.

{
  "PostToolUse": [{
    "matcher": "Write|Edit",
    "hooks": [{
      "type": "mcp_tool",
      "server": "slack",
      "tool": "post_message",
      "input": {
        "channel": "#dev-alerts",
        "text": "Claude edited a file"
      }
    }]
  }]
}

Installation checklist

chmod +x .claude/hooks/my-hook.sh   # Required
# Test manually before activating:
echo '{"tool_name":"Edit","tool_input":{"file_path":"test.ts"}}' \
  | .claude/hooks/my-hook.sh