Using ATL Sandbox

What ATL Sandbox does

ATL Sandbox gives ChatGPT and Claude isolated cloud computers for your account. Your assistant can create a sandbox, run code and shell commands in it, attach a GPU, keep files, use a private PostgreSQL database, publish what it builds, and schedule jobs, all from the conversation.

Every sandbox runs behind its own user-space kernel and only reaches the compute, storage, and database that belong to your signed-in account.

Connect your assistant

ChatGPT

1. In ChatGPT, open the apps directory and search for ATL Sandbox.
2. Select Connect.
3. Sign in or create an ATL account on the ATL sign-in page.

Claude

1. In Claude, open Settings → Connectors and browse the directory for ATL Sandbox.
2. Select Connect.
3. Sign in or create an ATL account on the ATL sign-in page.

If the listing is not available to you yet, add a custom connector with the server URL https://atl.vizpageai.com/mcp. The same account works in ChatGPT, Claude, and the web dashboard.

Things to ask

• “Create a Python sandbox and analyze this CSV. Which region grew fastest?”
• “Render my Blender scene on a GPU and save the frames to my storage.”
• “Build a landing page for my bakery and publish it.”
• “Every morning at 8, run report.py and save the result.”
• “Start a Node sandbox, write a small API, and run its tests.”

Your assistant picks the right tools, and usually asks you to confirm before it changes or deletes anything.

Sandboxes and environments

Ready-made environments: python, node, cpp (GCC), rust, and blender on CPU, plus cuda and blender-gpu with a GPU. Official Docker images such as python:3.13-slim also work. With the developer security profile, any public linux/amd64 container image that includes /bin/sh can be used.

• Instant sandboxes start in seconds on a disposable workspace. Stopping one discards its files.
• Durable sandboxes keep an encrypted persistent workspace across stops and starts.

GPUs

GPU sandboxes provide NVIDIA GPUs for CUDA, model inference, and Blender Cycles rendering. The l4-3gb slice starts in seconds; full GPUs can take a few minutes while capacity is added. Blender GPU sandboxes come preconfigured for CUDA rendering.

Files, storage, and databases

• Workspace: files inside a sandbox under /workspace.
• Object storage: private files for your account that outlive any sandbox. Your assistant can save workspace files there and read them back.
• PostgreSQL: every account has its own database, reachable only from your sandboxes.

You can browse sandboxes, storage, and files in the dashboard.

Publishing and automations

Your assistant can publish a static website from a sandbox, expose a running web app on a temporary public link, or deploy it as a separate app. Published content is visible to anyone with the link, so only publish what you mean to share.

Automations run a command in one of your sandboxes on a schedule (for example daily at 08:00) or once as a job, and keep a run history you can check later.

Plans, credit, and limits

Compute is prepaid and metered every 60 seconds. When an account's credit reaches zero, running compute stops; there are no overage charges. Plans and credit packs are managed in the dashboard, and plan details are on the pricing page.

The Free plan includes one concurrent sandbox with a one-hour run limit. Instant sandboxes, GPUs, persistent workspaces, and schedules are included in paid plans.

Security and privacy

• Each sandbox is isolated from the host and from other accounts.
• Storage and databases are scoped to your sign-in, and persistent storage is encrypted at rest.
• Infrastructure credentials never enter the conversation or your browser.
• Do not paste passwords, API keys, or payment details into the conversation or into sandboxes.

Read the Privacy Policy, Terms, and Acceptable Use Policy.

Troubleshooting

• The sandbox is still starting. First starts and GPU sandboxes can take a little longer; your assistant checks again until it is ready.
• “The sandbox failed to start.” A custom image must be linux/amd64 and include /bin/sh; distroless and scratch images cannot run.
• “Not included in this account.” The feature needs a paid plan; see the dashboard.
• “No compute credit left.” Add credit or wait for your plan's next monthly credit.