Willison publishes external sandbox prototypes for coding agents
The agent stays on your Mac, the code it writes runs in WebAssembly, QEMU or an Apple Container.
Simon Willison has described the external sandbox pattern, in which a coding agent runs on your Mac but executes the code it writes inside a WebAssembly, QEMU or Apple Container sandbox. He published demos for three agents, Claude Code, Codex and Pi, and an index repository collecting the prototypes.
a credential replacement mechanism to make authenticated API calls without sharing a secret with the agent
The work was done with Prime Radiant, which Willison describes as the AI research lab behind Superpowers.
What is actually in the box
There are three sandbox servers, one per sandboxing technology. Each one includes an optional HTTP/HTTPS allow-list proxy, so you can say exactly which hosts the code your agent writes is allowed to talk to. Each also ships a credential replacement mechanism, which Willison says lets the sandboxed code make authenticated API calls without the agent ever seeing the secret.
That second piece is the part worth stealing even if you never run any of these servers. Most setups today conflate two different permissions, the agent being able to call an API and the agent holding the key for that API. Credential replacement splits them. The proxy holds the secret and injects it, the agent gets the response, and a prompt injection or a confused tool call has nothing to exfiltrate.
Why external
The distinction in the name matters. The agent process itself is not sandboxed, it keeps running on your machine with your files and your editor. Only the code it produces gets shipped out to the isolated environment to run. That is a different trade from sandboxing the whole agent, which tends to break the parts of the workflow that make agents useful in the first place.
These are prototypes, not a product, and Willison frames them as exploration rather than a recommendation. But the shape of the thing is the interesting part. Allow-list the network, hold the credentials outside the agent, let the untrusted code run somewhere you can throw away. Nothing in that list requires a specific vendor or a specific model, which is probably why it is in a public repository rather than a release note.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.
