Skip to main content
Sandboxes provide isolated compute environments for running untrusted code, agent tools, evaluations, and reinforcement learning (RL) workloads. Create an environment, run commands, work with files, and collect results through the Python or TypeScript software development kit (SDK). Start on serverless capacity that CoreWeave operates, with no cluster to provision. When you need control over placement and policy, you can also run sandboxes on your own CoreWeave Kubernetes Service (CKS) clusters. Serverless sandboxes are generally available, and usage is billed. Access is enabled per organization. To request access for your organization, contact support.

​
Get started

Run your first sandbox

Authenticate and run a command with Python or TypeScript.

Run coding agents

Choose an integration for interactive coding or autonomous workers.

Run evaluations

Evaluate code patches, coding agents, and tool-calling models.

Build RL workflows

Run model-generated code in sandboxes for reinforcement learning reward evaluation.
Connect Claude Code, Claude Managed Agents, or Devin Outposts to sandbox workspaces. For your own application, explore the Python client and TypeScript SDK. Use sandboxes with Harbor, or explore the Sandboxes recipes for warm pool examples and the taubench-verl RL training recipe. The RL training guide uses the examples/rl_training/ scripts in cwsandbox-client. To run pull request tests in disposable sandboxes, see Run pull request checks in sandboxes.

​
Configure your sandboxes

Choose the features your workload needs:
  • Configure execution. Choose your image, CPU, memory, and lifetime with sandbox configuration, or create a multi-container sandbox from a supported Docker Compose file.
  • Save and reuse files. Capture a working directory on a scratch volume with file system snapshots, then restore it into new sandboxes.
  • Attach storage. Use volumes for scratch space. On CKS, registered volumes let sandboxes share existing storage.
  • Expose a service. Give a sandbox application a public address with public endpoints.
  • Use GPUs. Request GPU sandboxes on serverless or CKS capacity. GPU sandboxes require separate access. Use serverless GPU sandboxes for untrusted code. GPU sandboxes on your own CKS cluster aren’t virtualized and should run only trusted code.

​
Choose where sandboxes run

CoreWeave operates serverless capacity and sets its policy. Serverless is the default. Follow the serverless quickstart to choose a credential and start running sandboxes. CKS uses clusters in your organization. Your administrator enables a runner and configures its sandbox policy. For deployment and scheduling details, see Use your own compute, architecture, and CoreWeave SUNK integration.

​
Migrate to cwsandbox

If you use the deprecated wandb.sandbox module, follow Migrate from wandb.sandbox. The guide explains how to install the standalone Python client, retain your authentication choice, and review wrapper defaults.
Last modified on November 5, 2026