Get started with Docker Agentic Platform

Use a cloud sandbox to have an agent review a project without setting up the project on your computer. In this walkthrough, you'll launch Claude Code, clone a public sample application, and save a project review. Then you'll pause the sandbox to return later or delete it when you're finished.

You can complete the walkthrough in the browser. An optional step shows how to open the same project in VS Code or Cursor over SSH.

Before you begin

You need:

• A personal Docker account with an active Docker Agentic Platform subscription.
• An Anthropic API key for the Claude Code kit. Add it under Secrets or enter it when you launch the sandbox. Your model provider bills inference separately from Docker's sandbox compute charges.

This example uses a public repository, so you don't need a GitHub token. For other agents and their credentials, see Kits and Secrets.

Start a sandbox

1. Open Docker Agentic Platform and select New.
2. Choose the Claude Code kit.
3. Add your Anthropic API key if it isn't already saved. Don't enter it in the agent's prompt or in a file inside the sandbox.
4. Keep the defaults and select Run.

On your first launch, the default Open network policy permits outbound access to any host. If you're using a custom policy, make sure it permits GitHub and your model provider. See Network policies.

When the sandbox is ready, its terminal opens with Claude Code. After launch, you can't change the selected credentials, tools, network policies, or compute size. See Sandbox options for other launch settings.

Give the agent a task

Your local files aren't mounted in a cloud sandbox. Ask Claude Code to clone and review a public sample project:

Clone https://github.com/docker/welcome-to-docker.git into
/home/agent/workspace/project, then review the application.
Read the source, package.json, and Dockerfile. Don't modify the application
or install dependencies.
Write /home/agent/workspace/review.md with:
- A short explanation of what the application does.
- The commands defined in package.json for running and building it.
- An explanation of how the Dockerfile builds and serves the application.
- One small improvement you'd recommend, with the relevant file path.
Base the review on the files you read. Don't claim to have run the application.
Summarize your findings in your response.

Claude Code summarizes the project in the conversation and saves the review at /home/agent/workspace/review.md. You can ask follow-up questions or give it another task in the same session.

Keep your work

The cloned project and review are stored in the sandbox. Files don't sync automatically with your computer or a remote repository. They remain available when you pause and resume the sandbox, but are lost when you delete it.

To keep files outside the sandbox, use the CLI to transfer files, or commit and push changes to a repository. Include a GitHub credential when launching a sandbox that needs to push changes.

Open the project in your editor (optional)

Open the project and review in VS Code or Cursor over SSH. Your editor runs on your computer, while the files and terminal commands stay in the cloud sandbox.

1. Install the Docker Sandboxes CLI, version 0.45.0 or later.

2. In a terminal on your computer, sign in with the same Docker account you used in the Console and configure cloud SSH access:

   $ sbx login
   $ sbx --cloud setup ssh
   $ sbx --cloud ls
   

3. Find your sandbox in the list and copy its sbx_-prefixed ID. Test the connection, replacing <SANDBOX_ID> with that ID:

   $ ssh <SANDBOX_ID>@sbx_cloud
   

   A successful connection opens a shell inside the cloud sandbox. Run exit to return to your computer's terminal.

4. In VS Code, install the Remote - SSH extension. In Cursor, install its Remote - SSH support.

5. Open your editor's Command Palette and run Remote-SSH: Connect to Host. Enter <SANDBOX_ID>@sbx_cloud, using the same ID as in the SSH command.

6. Open /home/agent/workspace in the remote window. You'll find the cloned application in project and the agent's output in review.md.

The first connection installs the editor's server inside the sandbox.

To download the review, run this command in a terminal on your computer:

$ sbx --cloud cp <SANDBOX_ID>:/home/agent/workspace/review.md ./review.md

Pause, keep running, or delete the sandbox

Closing the browser or disconnecting your editor leaves the sandbox running. An agent can continue working until you pause or delete the sandbox, or its timer expires. Compute is billed while the sandbox runs, even when the agent has finished its task.

Return to Sandboxes in the Console and choose what to do next:

• Pause the sandbox to stop compute charges and keep its files. Confirm that it is paused in the list. To continue later, reopen the sandbox and resume it. Resuming starts a fresh timer with the original duration.
• Leave it running if the agent still has work to do. Check the expiration time: the default Stop action stops the sandbox after one hour, including background processes, and keeps its files.
• Delete it when you no longer need it. Save the review or push any changes you want to keep first. Deletion removes files stored only in the sandbox.

For more detail, see Lifecycle actions. To run a task from your own terminal, follow the cloud CLI walkthrough. For programmatic workflows, see the Sandboxes API quickstart.

Sandbox options

Configure these options before selecting Run:

The launcher remembers your policy selections from the previous launch in the same browser. Policies marked Always applied are selected automatically. Review the selected policies before launching. Selecting Balanced alongside Open doesn't restrict Open's access; see How policies combine.