Using WSLc as a sandbox for AI coding agents on Windows

[Abechan (abechan1)]

Using WSLc as a sandbox for AI coding agents on Windows

I've been working on wip, a CLI for Windows development environments, and recently started using WSLc as a sandbox for AI coding agents.

The motivation is simple:

AI coding agents such as Claude Code and Codex are much more useful when they can freely run commands, install dependencies, edit files, and execute tests.

But giving an agent broad permissions directly on the Windows host is something I wanted to avoid.

So I'm experimenting with this model:

Windows
↓
wip
↓
WSLc sandbox
↓
Claude Code / Codex / other coding agents
↓
repository

The idea is to create an isolated, disposable Linux environment for a repository while keeping the developer workflow on Windows.

With this approach, the agent can be given much broader permissions inside the sandbox without giving it equivalent access to the host machine.

The goal of "wip sandbox" is to make this workflow simple enough that the sandbox becomes the default place to run coding agents, rather than something developers have to manually set up each time.

Some of the areas I'm working on include:

• creating WSLc-based sandboxes from a simple CLI
• repository-specific isolated environments
• persistent volumes where persistence is useful
• disposable environments when isolation is more important
• support for multiple AI coding agents rather than coupling the environment to one agent
• keeping the Windows-side setup as small as possible

Repository:

https://github.com/slidict/wip

WSLc becoming generally available made this approach especially interesting, because Windows now has a lightweight container primitive that can be used directly for development tooling like this.

I'm still refining the sandbox workflow, so I'd be particularly interested in feedback from people experimenting with WSLc:

• Are you already using WSLc for development environments?
• Are there isolation or filesystem behaviors I should be careful about?
• Does using WSLc as an execution boundary for AI coding agents make sense to you?

I'd also be interested to hear about other WSLc-based developer tools being built.