Atlas: sandboxed AI workspaces

Context

Coding agents such as Claude Code and Codex can now make real changes to real websites. For a developer that is a large speed-up. For a business owner who isn't technical, it is tempting and risky: the same tool that can fix a typo can, unsupervised, reach anything else on the server.

The challenge

Let a non-technical owner change their own site by describing what they want, without giving an AI agent the keys to everything else. That means each client gets a workspace tailored to their site, the agent inside it can only reach that client's project, and nothing a client does can affect another client.

What I did

I built Atlas, a platform that runs Claude Code and Codex behind a tailored AI workspace for each client, with strict sandboxing.

• Per-client workspace. Each client works in a workspace set up for their own site and their own kind of changes.
• Strict sandboxing. The agent in a workspace can only reach that client's project; every other project on the server is out of reach.
• Built for non-technical owners. The owner asks for a change in plain language from inside their own site's admin area, and the change is made for them.
• In production. A coaching business uses Atlas through its admin AI panel today, fully isolated from every other project.

Stack

Claude Code and Codex, running in per-client sandboxed workspaces on Linux.

Outcome

Atlas is live. A coaching business makes changes to its own site through it, safely and without touching anything else on the server. For a commerce team the lesson carries straight over: I use AI as leverage every day, and I know how to fence it in.