Sandboxing (New Since the Video)
Instead of approving commands one by one, you can let the operating system enforce a boundary — which files and network domains Claude is allowed to touch.
The video's permission model is entirely about approval prompts: Claude Code asks, you answer. Sandboxing takes a different, complementary approach: you define which files and network domains a command is allowed to touch, and the operating system itself enforces that boundary for every shell command Claude Code runs (and anything that command spawns) — not Claude Code's own judgment.
This lets Claude run far more autonomously and safely at the same time: instead of being asked to approve every shell command, most commands just run inside the sandbox, and only things that would cross the boundary get flagged. It's built in and runs on macOS, Linux and WSL2 (native Windows isn't supported — run inside WSL2 there). Turn it on and configure it with /sandbox.
Real-life use cases
Letting Claude run tests unattended, safely
A developer enables the sandbox scoped to the project directory with no network access, then lets Claude iterate on a test suite for an hour unattended — confident that even a wildly wrong command can't touch files outside the project or phone home anywhere.
Fewer prompts without going “dangerous”
Instead of reaching for --dangerously-skip-permissions to stop constant interruptions, a user configures the sandbox with a defined file/network boundary — getting the speed benefit without removing the safety net entirely.
Source
- Configure the sandboxed Bash toolhttps://code.claude.com/docs/en/sandboxing