ATL Sandbox

By VizPageAI

Overview

ATL Sandbox gives ChatGPT isolated cloud computers for your account. Ask in plain language and ChatGPT creates a sandbox, runs Python, Node.js, C/C++, Rust, or shell commands, and returns the results to the conversation. What you can do: - Run code in ready-made environments (Python, Node.js, GCC, Rust, Blender) or, with the developer profile, any public Linux container image. - Attach NVIDIA GPUs for CUDA, model inference, and Blender Cycles rendering. - Keep work in encrypted persistent workspaces and private object storage, and use a PostgreSQL database that belongs to your account. - Publish static websites and expose web apps on public links when you choose to share them. - Schedule recurring jobs and one-off runs, and check their history later. Who it is for: developers, analysts, students, and creators who want their assistant to run real code and keep the results. Security: every sandbox runs behind its own user-space kernel and can reach only the compute, storage, and database of the signed-in account. Infrastructure credentials never enter the conversation. Actions that change or delete data are marked so ChatGPT asks before running them. Limitations: an ATL Sandbox account is required. Accounts have usage limits on concurrent sandboxes, run time, storage, GPUs, and schedules, and running compute stops when an account reaches its limit. Instant sandboxes discard their files when stopped. Images must be linux/amd64 and include /bin/sh. Interactive terminals and Codex sessions are not available in this app. Do not paste passwords, API keys, or payment details into sandboxes.

Tools

attach_sandbox_object_storage

ChatGPT

Give an existing sandbox access to the user's private object storage. Adds the storage gateway address and an expiring, storage-only token to the sandbox environment (SANDBOX_STORAGE_ENDPOINT and SANDBOX_STORAGE_TOKEN). A running sandbox restarts once to pick them up, which ends its running processes; an instant sandbox also loses the files in its disposable workspace. No cloud or infrastructure credentials are placed in the sandbox.

cancel_automation_run

ChatGPT

Request cancellation and terminate a running scheduled command on a best-effort basis.

cancel_sandbox_job

ChatGPT

Prevent a queued one-shot job from starting; cancel a running attempt by run ID instead.

create_gpu_sandbox

ChatGPT

Create an isolated one-GPU ATL sandbox. Call list_gpu_types and list_sandbox_environments first, then select gpu_type and a GPU image alias. This waits briefly, but GPU capacity can take several minutes to schedule when autoscaling starts from zero. If ready is false, keep calling wait_for_sandbox until it is true, then use run_python to verify torch.cuda or run_command with nvidia-smi. Use security_profile=developer when apt-get must install system packages; it adds only the package-manager capability set, and accepts any public image when developer_public_images is true. l4-3gb requires startup_class=instant and uses disposable storage with no checkpoint, automatic save, or restart after capacity loss. Lower-cost interruptible capacity is preferred with standard-capacity fallback.

create_persistent_storage

ChatGPT

Create encrypted persistent storage owned by the authenticated user.

create_sandbox

ChatGPT

Create a CPU sandbox using an administrator-approved environment image. Call list_sandbox_environments, then set image to an advertised CPU alias. Use restricted by default, or developer to permit apt-get with a limited capability set; when developer_public_images is listed as true, developer also accepts any public container image. Runtime system-package installs are ephemeral. Set memory_gib to reserve more memory; it is a fixed reservation, not a burst ceiling. startup_class=durable keeps the current encrypted persistent workspace. startup_class=instant uses a disposable workspace on platform-selected capacity, with no persistent attachment, checkpoint, automatic object save, or automatic restart after infrastructure loss. Scheduled jobs can recreate it and retry infrastructure failures. Use the provided tenant-scoped object storage gateway explicitly for any files that must survive. Creation waits briefly; if ready is false, call wait_for_sandbox until it is true.

create_sandbox_automation

ChatGPT

Create a restricted SandboxSchedule or desired-state SandboxController. Exec schedules may start a stopped sandbox, wait for readiness, and clean it up after the run. Durable sandboxes retain persistent storage; instant sandboxes delete their disposable runtime. Instant exec schedules retry infrastructure failures in a fresh sandbox up to max_attempts. Nonzero command exits are application results and are not retried. Infrastructure placement, identity, storage, network, and security fields are not accepted.

delete_deployed_app

ChatGPT

Delete a production copy and its private route without modifying the development sandbox.

delete_persistent_storage

ChatGPT

Permanently delete named persistent storage after exact confirmation; default storage cannot be deleted.

delete_sandbox

ChatGPT

Delete sandbox compute. Instant workspace data is discarded with its Pod.

delete_sandbox_automation

ChatGPT

Delete one tenant-owned virtual automation without deleting its sandbox.

delete_sandbox_job

ChatGPT

Delete a terminal one-shot job record; associated run history expires independently.

delete_sandbox_keep_storage

ChatGPT

Delete sandbox compute after confirmation while retaining independently managed persistent and object storage. Sandboxes can share named tenant storage, including the default workspace. Delete non-default storage only with delete_persistent_storage after all sandbox references have been removed. Default storage is never deleted.

deploy_sandbox_app

ChatGPT

Publish an app from a sandbox as a production copy on a public URL that anyone can open. Optionally runs build_command in the source sandbox first (it can change files there), packages source_path, and starts the app in a separate restricted sandbox. Remove the public deployment with delete_deployed_app. build_command runs from build_path, which defaults to source_path. Set build_path=/workspace when a build creates source_path as its output directory. Each production replica downloads the immutable release from tenant-isolated object storage into its disposable /workspace.

disable_sandbox_app

ChatGPT

Remove a temporary public route without stopping or modifying the development sandbox.

download_object_artifact

ChatGPT

Return one tenant-scoped object as an MCP file reference that ChatGPT can download. Use this after list_object_artifacts or save_workspace_file_to_object_storage when the user asks to receive a generated file in the conversation. The reference grants temporary read-only access to exactly one object.

download_workspace_artifact

ChatGPT

Return a workspace file as a downloadable ChatGPT artifact in one operation. Use this when the user asks to receive a generated file. workspace_path may be relative to /workspace or an absolute /workspace path. The file is copied to tenant-isolated object storage and exposed through a temporary, read-only reference; sandbox credentials and raw base64 are never returned.

expose_sandbox_app

ChatGPT

Temporarily expose one port of a running sandbox through a random public HTTPS hostname.

get_automation_run

ChatGPT

Read one tenant-owned automation run and its bounded output.

get_billing_account

ChatGPT

Return this authenticated user's plan, limits, and prepaid compute balance.

get_checkpoint_operation

ChatGPT

Poll an asynchronous checkpoint or restore request until phase is Completed or Failed.

get_command_result

ChatGPT

Poll a long-running run_command or run_python request without replaying it.

get_sandbox

ChatGPT

Inspect one sandbox without waiting. To wait for startup, call wait_for_sandbox instead.

get_user_database

ChatGPT

Return the authenticated tenant's PostgreSQL status without exposing database credentials.

list_automation_runs

ChatGPT

List bounded execution history for the authenticated tenant.

list_gpu_types

ChatGPT

List GPU models and memory sizes available for new GPU sandboxes. The response intentionally describes GPU capabilities without exposing backing infrastructure.

list_object_artifacts

ChatGPT

List this sandbox's artifacts in tenant-scoped object storage.

list_object_storage

ChatGPT

List the authenticated user's private object directory.

list_persistent_storage

ChatGPT

List only the authenticated user's persistent storage and attachment state.

list_public_apps

ChatGPT

List temporary previews and production copies owned by the authenticated user.

list_sandbox_automations

ChatGPT

List only the authenticated tenant's virtual schedules and controllers.

list_sandbox_environments

ChatGPT

List environment aliases and the image sources accepted by sandbox creation. Pass an alias, or any image reference from a trusted source (for example "python:3.13-slim" or "mathworks/matlab:r2026b"), as image to create_sandbox or create_gpu_sandbox. When developer_public_images is true, security_profile=developer also accepts any public image.

list_sandbox_jobs

ChatGPT

List the authenticated tenant's retained one-shot job records.

list_sandboxes

ChatGPT

List this user's ATL sandboxes and their lifecycle state.

publish_static_website

ChatGPT

Publish an object directory as an immutable public static website release. The source must contain index.html. Publishing copies only this tenant's bounded source directory to a separate private origin. Returned URLs always include index.html.

read_object_storage_text

ChatGPT

Read a bounded UTF-8 object from the authenticated user's private object directory.

read_object_text_artifact

ChatGPT

Read a UTF-8 artifact from this user's server-enforced object directory.

read_workspace_file

ChatGPT

Read a UTF-8 artifact from the user's persistent workspace.

register_user

ChatGPT

run_command

ChatGPT

Run a command. Long commands return a request_id that can be polled with get_command_result.

run_python

ChatGPT

Run Python. Long commands return a request_id that can be polled with get_command_result.

save_workspace_file_to_object_storage

ChatGPT

Copy a bounded non-secret file to object storage and return it to ChatGPT. workspace_path may be relative to /workspace or an absolute /workspace path. The returned temporary, read-only file reference avoids requiring a second download tool call. Never copy credentials or tokens.

scale_deployed_app

ChatGPT

Set the number of identical EBS-free production replicas serving an application.

set_sandbox_automation_enabled

ChatGPT

Pause or resume one tenant-owned automation using optimistic concurrency.

start_sandbox

ChatGPT

Start a stopped sandbox, waiting briefly for it. If ready is false, call wait_for_sandbox until it is.

stop_sandbox

ChatGPT

Stop compute. Instant workspace data is intentionally discarded.

submit_sandbox_job

ChatGPT

Submit an immediate one-shot command against an instant sandbox. The worker starts a fresh disposable Pod, executes asynchronously, retries only infrastructure failures, and deletes the Pod after completion. Poll list_automation_runs for output and status.

wait_for_sandbox

ChatGPT

Wait up to about 40 seconds for a starting sandbox to become ready. Use this after create or start returns ready=false. While the sandbox is still starting it returns ready=false with a next_action, and calling it again continues waiting. A GPU sandbox can take up to 10 minutes to start, which is expected rather than a failure.