Sandboxes

Sandboxes are short-lived, project-scoped, isolated compute environments. They’re built for ephemeral work — running untrusted code, evaluating a model, executing a one-off training job — without the lifecycle of a long-running pod.

A sandbox runs on a fraction or whole GPU, terminates on a timeout, and can be paused and resumed without losing state. Find it under Sandboxes in the sidebar. The feature is gated by REACT_APP_ENABLE_SANDBOX.

Sandbox vs. Pod

SandboxPod
Auto-terminates on a timeoutRuns until you stop it
Pause/resume preserves stateStop releases the GPU
Single image, single workloadLong-running, often multi-step
Use for: code exec, one-off jobs, evalsUse for: training, inference servers, notebooks

Creating a Sandbox

The wizard at Sandboxes > Create Sandbox walks through three steps.

1. GPU Selection

Pick a GPU type, count, and fraction:

GPUMemory
No GPU—
T416 GB
A1024 GB
L40 / L40S48 GB
A100 (40 GB / 80 GB)40 GB / 80 GB
H10080 GB

For each GPU you can take a fractional share (25%, 50%, 75%, or 100%) — useful when you need GPU access but don’t need a whole card.

2. Image

Choose one of the curated sandbox images (organized by category — PyTorch, vLLM, code interpreter, etc.) or provide a custom image URI. Custom images can pull from private registries by supplying a username and password.

3. Configuration

  • CPU and Memory (e.g. 1 CPU, 2Gi memory)
  • Timeout in seconds — the sandbox is terminated automatically after this duration (default 3600 s)
  • Environment variables — key/value pairs available inside the sandbox
  • Project — the sandbox is scoped to the selected project

Click Create Sandbox to launch. Provisioning typically takes seconds to a couple of minutes depending on image pull.

Lifecycle

StateDescription
RunningActively executing; billing is live
PausedState preserved, GPU released; billing reduced
TerminatedStopped — either by you, by timeout, or by failure; cannot be resumed
FailedProvisioning or runtime error; see logs

Actions on the sandbox list:

  • Pause — frees the GPU but keeps disk state; resume picks up where you left off
  • Resume — bring a paused sandbox back online (may schedule on a different GPU)
  • Delete — terminate and clean up

The sandbox list auto-refreshes every 10 seconds while there are active sandboxes.

KYC Requirement

Creating a sandbox requires KYC verification. Unverified users are prompted to complete KYC before provisioning.

Use Cases

  • Code execution in chat agents — run model-generated code against a real interpreter without trusting the model
  • One-off evals — fire up a sandbox, run a benchmark, capture metrics, tear down
  • Reproducing a bug report — bring up an isolated env matching the user’s stack
  • Short training experiments — when you’d otherwise leave a pod idle between iterations

Billing

Sandboxes are billed per second on GPU/CPU/memory while Running. Paused sandboxes accrue a much smaller storage charge for the held state. See the wallet expenditure breakdown for current rates.

Next Steps