Launch Codes

What it does

When Claude tries to run a dangerous shell command, it stops and a red-alert pane opens on the right: moving hazard stripes, the command, why it's dangerous and a four-letter launch code that expires in 30 seconds. A siren plays until you answer. The command only runs if you type the code and then choose LAUNCH. Anything else blocks it, and Claude is told not to retry or work around the check.

It asks for codes before:

• rm -rf, except on throwaway folders like node_modules, dist, .next, coverage and /tmp
• git push --force (also -f and +branch) and git reset --hard
• DROP or TRUNCATE sent through a SQL client such as psql, mysql or sqlite3, and supabase db reset
• vercel --prod, chmod -R 777, and a script from curl or wget piped into a shell

How to use it

• When the alert is up, a dialog under the conversation offers ABORT, ABORT and find a safer way, or Type something. To let the command run, pick Type something, enter the code from the pane, then choose LAUNCH.
• /launch-codes demo runs a practice alert with a made-up command. Nothing runs.
• /launch-codes test <command> tells you whether a command would need launch codes, without running it.

Good to know

• It only checks Bash commands, and it reads the command text, so it's a speed bump rather than a security boundary.
• If the code expires before you type it, the command is blocked.
• It has no network, file or process access. It draws the pane and plays the siren and launch sounds.