Simon Willison has published three open-source sandbox servers for code run by AI agents, using different isolation approaches while sharing the same local API. The design aims to let an agent run untrusted code without giving that code the keys to the machine hosting the agent.
Simon Willison Watch analysis
What happened
Willison says the project was developed with Prime Radiant. In his blog post, he describes an arrangement where the agent stays outside a dedicated sandbox and sends code-execution requests into it.
The three implementations use WebAssembly through v86, Apple Containers, and Microsoft’s Quicksand sandboxing library. Each offers the same localhost API, which can be configured for tools including Pi, Claude Code and Codex. An optional HTTP/HTTPS allow-list proxy restricts which hosts executed code can contact; a credential-replacement mechanism supports authenticated API calls without handing secrets to the agent.
Why it matters
An agent may need to run code to test a change, but code from a prompt, repository or dependency can be unsafe. Separating execution from the agent offers developers a way to experiment with different isolation setups without treating a promising coding assistant as a security boundary in its own right.
The shared API is the useful bit for anyone comparing approaches: the agent-facing interface can stay the same while the sandbox implementation changes. That is a more practical starting point than arguing about which single isolation technology should win the internet.
Our read
This is a useful pattern, not a magic safety seal. The allow-list and credential handling address two real ways executed code can cause trouble, while the choice of three implementations gives technically minded teams room to test what fits their setup. The next question is how well each option holds up under real workloads and deliberate attempts to break out.
What to watch
- Whether the three implementations receive maintenance and security review.
- How the sandboxes handle attempts to reach blocked hosts or escape isolation.
- Whether agent tools beyond those named by Willison adopt the shared local API.
Discussion spark: For code run by AI agents, would you trust one well-maintained sandbox implementation, or is the ability to compare different isolation approaches more important?
Sources and evidence
- External sandboxes for coding agents (17 November 2026, 22:52 UTC)
Independent WittyWires tracker for public updates about Simon Willison. Not affiliated with or endorsed by Simon Willison; this is not an official account.