What happened
Prime Radiant’s announcement describes “external sandboxes”: the coding agent runs on the user’s machine, while commands it executes are sent to a separate Linux environment. The company released three compatible prototypes: external-sandbox-v86, a Node.js process that runs Alpine Linux using the v86 x86 emulator compiled to WebAssembly; external-sandbox-container, a Python process that wraps Apple Containers; and external-sandbox-quicksand, a Python process that runs a QEMU virtual machine using Microsoft’s Quicksand. The announcement identifies them as prototypes, not production-hardened products.
Each implementation exposes the same localhost web API, and Prime Radiant also published a protocol specification and conformance-testing tool. The v86 server is distributed through npm and can be started with npx; the two Python servers are published on PyPI and run with uvx. The Apple Containers and Quicksand examples target Apple silicon Macs. The v86 example requires Node.js 22.19 or later. A Pi demonstration starts the sandbox and agent in separate terminals using two npx commands; the page also documents setup for Claude Code, Codex, and Pi. Pi’s extension requires version 1.5 or later. Prime Radiant reports that its automated Codex tests passed against Codex CLI 0.164.0 using a simulated model, while noting that Codex’s exec-server protocol is experimental and other versions may need checking.
The sandbox sees only folders that the user mounts, with read-only and read-write mounts available. The examples mount the current project directory read-write. Network access is disabled by default. Builders can enable HTTP and HTTPS through a proxy that restricts the allowed hosts and HTTP methods. Its credential-replacement feature substitutes a secret held by the host-side sandbox server for a placeholder token in requests, so the key itself is not placed inside the sandbox or shown to the agent. A web interface provides an in-browser terminal, network-rule controls, and a log of allowed and denied requests.
The agent integrations do not constrain every operation equally. Claude Code sends Bash commands through the sandbox, but its native Read and Edit tools still access the host. Codex sends process and filesystem operations, including apply_patch, through its integration. The Pi extension replaces bash, read, write, edit, grep, find, and ls tools; its documented --no-builtin-tools option prevents fallback to Pi’s local tools if the extension fails to load. Prime Radiant says the shared protocol is intended to let different agents and sandbox implementations be mixed and matched.
What the archive adds
New Runtime’s earlier coverage described the runtime boundary as a critical security control for agents that install packages, execute commands, and access files (earlier note, 2026-10-21). It also covered Microsoft’s Quicksand Python API for QEMU sandboxes (earlier note, 2026-10-24). This is a new release on a known story: Quicksand now appears as one backend in a compatible set of external sandboxes for coding agents.
Trend signal
This supports the view that agent security depends on runtime boundaries, rather than relying only on instructions or a model’s judgment. Prime Radiant’s examples make the boundary concrete through mounts, network policy, and host-side credential substitution. The announcement does not establish how robust those controls are under attack; the prototypes remain experimental. See Agent security moves to runtime boundaries.
Why now
Prime Radiant says the work grew from exploring safer ways to run coding agents, including with smaller, less capable models. The announcement argues that less capable models are more prone to mistakes and more vulnerable to deliberate prompt injection. Its proposed alternative to approving every action is to constrain what commands can access and where they can connect.
Connections to other events
New Runtime’s earlier coverage of Akamai’s beta launch paired managed model serving with an agent sandbox (earlier note, 2026-10-31). Prime Radiant’s release is a different approach: local agents connect to one of three compatible sandbox servers, rather than the announcement describing a managed cloud offering. Both put execution isolation on the product map, but this release makes the agent-to-sandbox protocol and its policies especially visible.
Why it matters
A coding agent’s permissions are not just a setting in its prompt. If an agent can read a credential file, alter unrelated projects, or send arbitrary network requests, a mistake or prompt injection can turn those capabilities into damage. External execution can narrow the command environment without moving the agent itself, and the shared API offers a possible way to switch between sandbox technologies without changing the basic pattern.
The important detail is that “sandboxed” does not mean every agent action is contained. Claude Code’s native file-reading and editing tools remain on the host in this integration, while mounted read-write folders are still exposed to damage. The network proxy and mount configuration therefore need to be matched to the actual agent integration and task. These prototypes provide a starting point for testing those boundaries, not evidence that teams can safely hand them sensitive workloads.
Limits and uncertainty
Prime Radiant explicitly calls the implementations experimental prototypes and says they have not been audited to a level suitable for sensitive workloads. The announcement gives no independent security evaluation, benchmark of isolation strength, or measured performance comparison among v86, Apple Containers, and Quicksand. It also says Codex’s exec-server protocol is experimental, so compatibility with other CLI versions is not established by the reported tests.
Isolation varies by agent: Claude Code’s native Read and Edit tools remain outside the sandbox. A read-write mount can still be damaged by the agent, and Linux environments may require dependencies built for macOS to be reinstalled. The proxy’s credential replacement also cannot prevent an allowed service from echoing a credential back in its response, so the authors recommend trusting services and narrowing methods and paths.
Skill to build now
For a low-risk trial, start with one disposable, version-controlled project and mount only that directory. Make any additional directories read-only, leave networking disabled until the task requires it, then allow only the specific hosts and HTTP methods needed through the proxy. If using Pi, disable built-in tools as the announcement’s example does; if using Claude Code, account for native Read and Edit operations that remain on the host. Confirm execution by running !uname -a in Pi and checking that it reports Linux.
Reader takeaway
Prime Radiant’s release makes external sandboxing testable across three Linux backends and three coding agents, with a common protocol and concrete controls for mounts and network access. Treat it as an experiment, not a security guarantee: map which tools actually cross the boundary, keep the writable surface small, and do not use it for sensitive workloads until its security has been independently assessed.