sandbox-sdk

sandbox-sdk is Cloudflare’s official skill for applications that operate isolated Linux environments on Cloudflare Containers from Workers through Sandbox SDK. The canonical source is in the official cloudflare/skills repository at skills/sandbox-stable. According to the provider, this stable skill is for the regular @cloudflare/sandbox package. It is a working contract and a retrieval guide for compatible coding agents, not a standalone container service, not a substitute for a Cloudflare account, and not a promise that arbitrary code becomes safe automatically.

Purpose and product boundary

The skill helps agents review, plan, and maintain Sandbox applications. Cloudflare describes Sandbox SDK as an interface that lets applications execute commands, manage files, run processes, and expose services inside isolated containers. That fits agents that need controlled execution for untrusted or user-generated code, interactive development environments, data analysis, and automated verification. The product boundary matters: the repository supplies agent instructions, while Workers, Containers, and Sandbox SDK provide the actual platform. This entry therefore does not claim that the instructions replace a runtime library.

Version line and workflow

The primary source requires checking the package line before changing code. A project using the stable package with a matching stable container image belongs to this skill. A project using @cloudflare/sandbox@next or a preview container belongs to the separate sandbox-next skill. Moving between the two lines is a distinct task and should not happen incidentally, because their protocols and API assumptions can differ. For stable applications, the skill explains that sandbox.exec accepts a command string and waits for the command to finish, returning buffered results. Longer-running or streaming work must use the stable process APIs intended for those cases.

Security and data responsibility

Cloudflare documents isolated environments, resource controls, and protective mechanisms, but no platform makes unchecked input harmless. According to the provider, user and file content must be validated before it is placed into commands, paths, or environment values. Depending on the application, a separate sandbox per user or job can help prevent accidental sharing of files and processes. Network access, external services, container images, and persistent data require deliberate review. The skill does not replace threat modeling, privacy review, access control, logging, or human approval. Files and output may be processed by the selected agent, its model provider, and Cloudflare, depending on configuration and workflow. Secrets belong only in protected runtime configuration and never in skill text, logs, or version control.

Professional context and limits

For a responsible rollout, teams should compare the current Cloudflare documentation, installed type definitions, and concrete container configuration first. The official documentation covers architecture, lifecycle, API, options, and the security model; it is authoritative for current platform behavior. This catalog profile remains useful because it documents a versioned, independent SKILL.md with a clear stable product boundary. It states no fixed prices and does not replace checking current provider terms. The license statement refers to the Apache-2.0 license in the official Cloudflare skills repository. Teams should review package, image, and binding changes together and protect each execution with suitable tests and operational observation.