External sandboxes bring coding-agent code execution outside the agent
Simon Willison and Prime Radiant introduced an external sandbox pattern that keeps the coding agent outside a dedicated environment while running agent-executed code inside it. They released three server implementations: external-sandbox-v86 using Alpine Linux in WebAssembly, external-sandbox-container using Apple Containers, and external-sandbox-quicksand using Microsoft's Quicksand library. The servers expose a common localhost API that can connect to Pi, Claude Code, or Codex, with optional network allow-listing and credential replacement.
Simon Willison's Weblog reports that the project gives coding agents an isolated execution boundary and gives users controls over network access and secrets, directly addressing risks from agent mistakes and prompt injection.
External sandboxes for coding agents
Simon Willison’s Weblog
Sponsored by: Dynatrace — Your coding agent can read your code but can’t see how it runs in production. Read the blog to see how Dynatrace MCP closes that gap.
17th November 2026 - Link Blog
External sandboxes for coding agents. I've been exploring new ways to sandbox code that is executed by agents, with a focus on making it safer to try out local models without worrying that a mistake or prompt injection attack might cause them to take damaging actions on a machine.
#Prime Radiant#external-sandbox-v86#external-sandbox-container#external-sandbox-quicksand#v86#Alpine Linux#WebAssembly#Apple Containers#Quicksand#Pi#Claude Code#Codex