Safe Agent Sandbox -- Let It Work Without Letting It Roam

An agent should earn broader access by proving the narrow job first.

Safe Agent Sandbox designs a bounded place to run AI tools without handing them your home directory, production credentials, unrestricted networking, or an open-ended provider bill. It starts from the task and grants only the capabilities that task can justify.

What you get

• A boundary map covering files, commands, network destinations, ports, providers, and persistence
• A least-privilege capability table with allow, deny, and approval-required decisions
• Guidance for disposable directories, containers, isolated Git worktrees, read-only mounts, and synthetic fixtures
• Wall-clock, CPU, memory, disk, request, iteration, and spending limits
• A scoped credential plan that keeps secrets out of prompts, logs, commits, and deliverables
• Evidence capture for commands, exits, changes, network attempts, and artifacts
• Explicit stop conditions and a teardown checklist
• An allow, allow_with_limits, or deny decision for the proposed run

The skill does not pretend that a container is automatically secure, weaken host protections to make a demo pass, or silently move work to a paid provider. Real publishing, messaging, purchasing, deployment, and destructive actions remain separately authorized.

Use it before evaluating unfamiliar agents, running autonomous coding tools, or exposing a workflow to sensitive repositories and services.