Agents Sandbox — AI skill for Claude Code

Agents Sandbox

**Full power for your agents. Full safety for your machine.**

The best runtime environment for your AI agents.

The Permission Dilemma of Running AI Agents Locally

Every AI agent running on a personal machine today faces the same lose-lose trade-off. If you use Claude Code or Codex, you know the drill: the agent is midway through a task, and a dozen permission dialogs pop up. You click `Yes` until your fingers go numb. Or you switch to `Yes always` and spend the rest of the day wondering if the agent just `rm -rf`'d something important.

Claude Code and Codex both offer sandbox-like capabilities that let you configure permission levels, but in practice they boil down to two choices:

This is exactly why Mac Minis took off in the AI agent community. People started buying a dedicated machine just for the agent — that way, even if it trashes something, it's the agent's files, not yours. Makes sense — except now you own a whole extra computer.

Give Claude Code full permissions (`claude --dangerously-skip-permissions`) and you're greeted with Anthropic's own disclaimer: only run this in a sandbox or VM, because dangerous commands will execute without asking. The default option is to decline and exit. You can ignore the warning and run it on bare metal, but if something breaks, that's on you.

What Matters Most When Running AI Agents Locally?

Related Skills