sandbox-platform is the operational safety guide for Sandbox clients. It helps an agent choose the CLI or MCP surface, classify a workload before creation, observe asynchronous work, and finish by cleaning up the resulting environment.

Start by discovering what the deployment can do. Check the controller and inspect existing sandboxes only when reuse is explicitly allowed. Then classify repository trust, generated code, secrets, network, resources, and TTL before creating anything.

The skill directs clients to default network access to deny, prefer restricted over open, request only necessary resources, and use isolation: auto unless an explicit microVM requirement exists. A client cannot weaken a server isolation decision to obtain capacity.

Create, wait, exec argv, inspect, and delete. The skill treats an operation timeout as ambiguous and requires inspection before repeating a mutation.

1234sandbox create --tenant "$TENANT" --image "$IMAGE" --network restricted
sandbox wait "$OPERATION_ID" --timeout 900
sandbox exec "$SANDBOX_ID" -- cargo test --workspace
sandbox delete "$SANDBOX_ID"

Output has to be interpreted, not just displayed. Check the operation state, command exit code, stderr, and truncated field before declaring a remote command successful.

A returned tunnel URL is always public. For a local machine, use sandbox http PORT and keep it attached. For a managed Sandbox, bind the service to 0.0.0.0, create one HTTP/WebSocket tunnel, and delete it when done.

The skill rejects treating a hostname as access control. Read the tunnel guidance and Sandbox security before exposing a service.

The hosted Sandbox llms.txt index is the canonical discovery file. It lets an agent retrieve only the narrow architecture, API, CLI, MCP, safety, or operations reference needed for the task.

Open the hosted raw index, the LLM discovery file, or the Sandbox MCP guide as needed.