Concept illustration of a developer using a separate test workspace beside a main workstation.

AI Coding Sandboxes: Restrict Shell Access, Then Review the Code

A working prototype is only a starting point

The accessible introduction to heise’s coding guide contrasts one-prompt prototypes with ongoing software development. A downloaded application may appear to work while leaving architecture, maintenance and later changes unresolved. The full tutorial is behind a subscription; this article does not claim to reproduce or validate its setup.

What a sandbox can restrict

Claude Code’s documentation describes operating-system restrictions around shell commands and their child processes, covering filesystem access and network destinations. Those boundaries reduce what an agent can reach while executing commands. They are separate from the question of whether the resulting application behaves correctly.

Check which tools the boundary covers

The documented shell sandbox does not cover every tool: file tools, MCP servers and hooks run outside it. Support also depends on the platform. The documentation lists macOS, Linux and WSL2, while native Windows commands run without that sandbox. Treating the word “sandbox” as a blanket guarantee would hide these distinctions.

Practical interpretation and limits

Keep execution boundaries and software review as separate checks. This is an explanation based on the accessible introduction and product documentation, not a tested installation guide or proof that an agent cannot escape its permissions.

Sources

Privacy Overview

This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.