Skip to main content
A sandbox is an isolated runtime an agent can be trusted with: run commands, read and write files, drive a desktop, open a tunnel, call an LLM through a key-holding proxy — all inside a container that dies on schedule.

​
Where to start

Create and execute

A sandbox with CPUs, memory, isolation and a timeout — then run commands in it.

The end-to-end guide

Create → execute → files → lifetime → teardown, with the cost-capped session pattern.

Files

Write, read, list, glob, move, copy, watch — the full toolkit.

Secrets & inference

Store provider keys once; the inference proxy injects them so keys never enter the sandbox.

​
The surface, by group

The Computer Use engine drives full jobs against real tools; the Sandbox is the lower-level runtime — the place an agent’s own code and desktop actions execute in isolation. Computer Use runs on sandboxes; you can too.
  • Sandboxes — concepts, architecture, deployment
  • Agent sessions — long-running loops with a cost cap and rotation
  • Webhooks — the 24 events and HMAC verification
Last modified on August 31, 2026