External sandboxes for coding agents
External sandboxes for coding agents
The post introduces the external sandbox pattern, in which code executed by an agent runs inside a dedicated sandbox while the agent itself lives outside it, aimed at making it safer to try local models against mistakes or prompt injection. Three sandbox server implementations were released with the same localhost API: external-sandbox-v86 (Node.js server running Alpine Linux in WebAssembly via v86), external-sandbox-container (Python web server running Alpine Linux in an Apple Containers instance), and external-sandbox-quicksand (Python web server using Microsoft's Quicksand library, bundling QEMU and Alpine Linux). Pi, Claude Code, or Codex can be configured to send API calls to the sandbox for code execution, and the sandboxes include an optional HTTP/HTTPS allow-list proxy plus a credential replacement mechanism for authenticated API calls without sharing a secret with the agent.
Source: Simon Willison · simonwillison.net