Simon Willison releases external sandboxes: run Claude Code/Codex/Pi inside a locked-down Linux VM
rickasaurus · x · 2026-11-17
Simon Willison (Prime Radiant blog) released the "external sandbox" pattern with three prototype implementations — Node.js + WebAssembly (v86 running Alpine Linux), Apple Containers, and Microsoft Quicksand with QEMU — all speaking the same protocol, letting Claude Code, Codex, or Pi execute every command inside a Linux sandbox that only sees the folders and hosts you allow.
- The simplest demo is two npx commands: start external-sandbox-v86 (needs Node.js ≥22.19) with the project mounted read-write, then launch the Pi coding agent with a sandbox-url; !uname -a confirms commands run in Alpine Linux.
- The agent itself stays on your Mac; only its executed commands are confined to the sandbox, unable to touch the rest of the machine.
- The motivation includes safely running smaller, less capable coding models; Claude Code and Codex need a bit more setup, with instructions in the post. rickasaurus amplified it with a joke about putting an agent in a Smalltalk VM.
More from coding & agent
- Startup security checklist: SSO/MFA, secret rotation, least privilege and more — Kyrannio · 2026-11-17
- Interactive visual app for learning system design, built for the vibe coding era — jakedahn · 2026-11-17
- Dan Shipper's 'proof of work': rank feeds by how many tokens agents spend on content — danshipper · 2026-11-17
- Ought launches Epistemics Fund: $500k–$5M grants for fact-checking and forecasting tools — NathanpmYoung · 2026-11-17
- Trading agents you create with one sentence: every decision logged publicly — bigrceo · 2026-11-17
- Building Agent-im: A Four-Layer Model for Agent Versioning, Memory and Secrets — yangyi · 2026-11-17