What problem does it solve? Running untrusted, user-submitted, or AI-generated code safely is hard: in-process sandboxes like vm2 have known escapes, and child_process or eval shares your host process. This Skill provisions isolated Firecracker microVMs on demand so risky code executes in a separate Linux VM with its own filesystem, network policy, and lifecycle. ## Core Features & Use Cases - Isolated code execution: Create a VM, run commands, read and write files, expose ports with public preview URLs, and stop or snapshot the environment when finished. - Persistence and snapshots: Persistent sandboxes auto-snapshot their filesystem on stop and resume later; manual snapshots let you boot new VMs with dependencies pre-installed. - Security controls for untrusted code: Apply deny-all or allow-list network policies, broker credentials at the egress firewall so secrets never enter the VM, and isolate multiple agents as separate Linux users. - Use Case: An AI coding agent needs to install dependencies and run tests on generated code. Create a sandbox with a 30-second timeout, vcpus: 1, and a deny-all network policy, run the test suite inside, then stop the VM in a finally block. ## Quick Start Ask the AI to create a Vercel Sandbox, run a Python command inside it, print the output, and stop the sandbox when finished.
