2026-11-17 · ← News
The coding agent stays on the Mac while its commands run in isolated Linux
Prime Radiant and Simon Willison released three external sandbox prototypes for Claude Code, Codex and Pi. The agent remains on the host, while its commands receive only selected folders, controlled network access and placeholder credentials.
The image could not be loaded.
Prime Radiant and Simon Willison have published a shared protocol and three experimental servers that separate a coding agent from the environment where its commands execute. Claude Code, Codex or Pi continues to run on the Mac. Depending on the client, shell commands, file operations and other supported tools are sent to Linux running in v86, Apple Containers or QEMU through Microsoft Quicksand.
Three sandboxes expose one local interface
All three implementations offer the same localhost API. According to its repository, the v86 package is about 36 MB including Linux. Users mount only selected directories and mark each one read-only or read-write. Networking is disabled by default. An optional HTTP and HTTPS proxy allows specific destinations and methods.
The proxy can replace a placeholder token with the real API key on the host. The secret therefore stays out of the sandbox and the agent context. The authors also warn that a trusted allowed service could transform and echo a credential, so rules should remain narrow.
Safety moves from model judgment into system boundaries
The practical change is that a bad command or prompt injection does not need another model to stop it. Filesystem and network boundaries do that job. This is particularly useful when testing cheaper or local models without relying on a vendor safety classifier.
The common protocol also separates agent integration from the virtualization layer. A team can compare WebAssembly, a container and a full virtual machine without rewriting the client. Security policy becomes part of the workflow rather than a stream of approval dialogs that people eventually click through.
The sandbox protects only operations that cross its boundary
The prototypes have not been audited for sensitive workloads. An agent can still damage anything in a read-write mount, and the Linux guest may need fresh dependencies instead of packages built for macOS.
Coverage also differs by client. Claude Code sends Bash into the sandbox, while its native Read and Edit tools remain on the host. The authors explicitly avoid claiming that every agent operation is isolated. This protects routed execution, not the entire product by default.
Audits and complete tool coverage will decide whether teams can trust it
The next test is an independent security audit and evidence that uncovered client features cannot bypass mounts or network policy. Compatibility with future Codex releases also matters because its exec-server protocol remains experimental.
Production teams should watch less glamorous measures: how many tools really execute inside the boundary, which directories remain writable and whether the proxy log captures every attempted connection. Those details determine whether a prototype becomes a dependable control.
Lilith's verdict
The agent can stay on the Mac while its hands work behind glass. That glass earns its value only when a prompt injection reaches for a file and hits a real wall.