Quick start¶
Five minutes from install to first sandboxed AI run.
1. Install¶
(See Installation for prerequisites and alternatives.)
2. Sandbox an AI agent¶
# cd into your project - this directory becomes the sandbox root
cd ~/projects/my-app
# Run Claude Code inside the sandbox
devsandbox claude --dangerously-skip-permissions
devsandbox wraps any command. Everything after the binary name is passed through to the sandboxed program. The flag --dangerously-skip-permissions is a Claude Code flag that disables its permission prompts - safe to enable here because devsandbox provides the actual security boundary.
3. Verify what's protected¶
This prints the sandbox configuration: which directories are mounted read-only, which are blocked, what network mode is active.
What just happened¶
~/projects/my-app(your CWD) became the project root with full read/write access.~/.ssh,~/.aws,~/.azure,~/.gcloud→ not mounted, invisible..envand.env.*files → masked with/dev/null..git/→ mounted read-only (no commits, no credentials).- mise-managed tools, your shell config, editor setup → mounted read-only so they work inside.
- Network → full access by default; add
--proxyto log every HTTP request.
Other tools, same pattern¶
devsandbox wraps anything with a CLI:
Next step¶
Continue to First run for a guided walkthrough that shows what the sandbox actually does.