Claude Code Hooks is AMAZING: "Text Message Me When AI Agent is Done"

Cloud Code’s new “hooks” feature lets developers attach custom shell-command automation to key moments in an AI agent’s workflow—so tasks can trigger real-world notifications, tests, and even Git operations without manual babysitting. In the most practical demo, a “stop” hook fires whenever Claude Code finishes, sending a Twilio text message to a phone. To prove it works end-to-end, the creator runs an agent to generate a small ping-pong simulator web project; once the agent completes, the phone receives a message confirming the task is done. For long-running agent work, that kind of push notification matters because it turns “check back later” into “get alerted immediately when something changes.”

The hook system also supports “post tool use” automation, enabling immediate follow-up actions after the agent calls a tool. One example generates unit tests automatically after Python code is produced: after creating a simple prime-checking tool, the agent writes a corresponding test file (test.py) and the workflow verifies correct behavior for prime inputs like 2 and 17 and non-primes like 1 and 4. The point isn’t that every project should auto-generate tests this way, but that hooks make it possible to chain quality checks directly into the agent’s output loop.

Next comes a more operational use case: using hooks to stage and commit changes to a GitHub repository. Because Cloud Code lacks checkpoints, the creator sets up a hook that runs shell commands to automatically `git add` changes and push them. After restarting Cloud Code to ensure the hook configuration in settings.json takes effect, the agent modifies prime-checking code; the repo status shows files staged and ready to commit. The workflow then expands to multiple iterations—adding a README, updating its author line, and refreshing the repository to confirm new commits appear. The demo even includes a “revert” style idea: using a hook-driven workflow to roll the repo back to an earlier commit, effectively acting like an automated checkpoint.

Finally, hooks can trigger multimedia feedback. A “post tool use” hook plays an MP3 sound from 11 Labs whenever the agent completes a tool call—so the creator hears a cat meow after running a small Python update. The overall takeaway is that hooks turn Claude Code from a purely interactive coding assistant into an automation layer: notifications for completion, tests for correctness, Git for version control hygiene, and audio/other side effects for immediate, human-friendly feedback.