Hooks

Claude Code hooks are user-defined shell commands (or HTTP requests, MCP tool calls, or prompt-based evaluations) that execute automatically at specific points in Claude Code's lifecycle. Unlike asking Claude to perform an action, hooks guarantee deterministic behavior — they always run regardless of what the model decides. This makes hooks suitable for enforcing policies, automating repetitive tasks, and integrating Claude Code into existing development workflows. Hooks are configured as JSON objects in Claude Code's settings files and can target specific lifecycle events such as before a tool runs (PreToolUse), after a tool completes (PostToolUse), when a session starts (SessionStart), or when Claude needs user input (Notification). Each hook definition specifies a trigger event, an optional matcher pattern to filter which tools trigger it, and a handler command to execute. The hook receives contextual data via stdin as JSON (for example, the tool name and its inputs), and can communicate back to Claude Code via exit codes and stdout/stderr output. A hook can silently allow an action, block it with an explanatory message, or inject additional context into Claude's next turn — giving developers fine-grained, programmatic control over the agent's behavior.