Docker Cloud Sandboxes move AI coding agents from laptop to cloud in one command

Docker has launched Cloud Sandboxes, a service that extends the isolated environments it built for AI coding agents from developers' computers to Docker-managed cloud servers, iThome and Docker's own blog report. Developers can let an agent carry out long tasks such as large refactors or test runs in the cloud, and move a sandbox between their laptop and the cloud with the sbx move command.

A sandbox, here, is an isolated environment in which an AI agent can install packages, run commands and start containers without reaching resources on the developer's machine that were not deliberately opened to it. Docker Sandboxes, launched earlier this year, run each agent in a microVM with its own Linux kernel, file system, network and Docker daemon, iThome explains.

Why agents need a place to run

The most important change in coding agents over the past year is that they can work for much longer, Docker writes: tasks that once needed a developer to check in every few minutes, such as a dependency migration or a test suite that takes an hour, can now be handed off and reviewed when they finish. A laptop is built around a person, the company argues: it sleeps when the lid closes, slows down on battery and disconnects when you move, which does not matter for a 30-second task but matters for one that runs all night.

Cloud Sandboxes use the same microVM isolation and the same command-line tool as the local version, according to Docker. Developers can start an agent in the cloud before leaving for the day and review its work in the morning, start locally and move a task to the cloud when it outgrows the laptop, or run many tasks in parallel, each in its own microVM with its own secrets and network policy.

What is included

• Kits: pre-built sandboxes for Claude Code, Codex, GitHub Copilot, Google Antigravity, OpenCode and Hermes, or a custom environment.
• MCP: connect servers such as Jira, Linear, Grafana or incident.io once, through a single gateway.
• Secrets: keys are stored once and injected per request, so the agent never sees the actual secret.
• Policies: network rules defining which endpoints agents can reach.
• Pricing: from 0.07 dollars an hour for 1 vCPU and 2 GiB to 1.12 dollars an hour for 16 vCPUs and 32 GiB, billed by the second.

The secret-handling design matters for security. The Cloud Sandboxes proxy injects credentials into each request, so agents do not see the actual secret; prompt injections can't touch secrets your agents never had in the first place, Docker writes. A prompt injection is an attack in which malicious instructions hidden in content the agent reads try to make it act against its user. Docker says centralised governance for enterprises is coming soon through a product called Docker AI Governance.

The limits of moving a sandbox

Moving between laptop and cloud does not preserve running processes, iThome notes. The sbx move command saves the sandbox's file system as a container image and creates a new sandbox at the destination, so running programs and memory contents are not carried over. Locally mounted directories, keys managed by sbx and network access rules do not move either; the destination needs its own configuration, and the original sandbox is kept. Docker itself warns that local and cloud sandboxes keep separate secrets, templates and network policies.

Cloud Sandboxes are pay-as-you-go: compute is metered by the second, a paused sandbox costs nothing, and volumes, outbound traffic and public images and kits are free, according to Docker. Model costs are billed separately by the model provider, iThome adds, and customers can bring their own key. Sandboxes run for one hour by default and up to 24 hours per session. The service requires sbx 0.45.1 or later and the pay-as-you-go plan on Docker Personal or Pro accounts; local Docker Sandboxes remain free and do not require Docker Desktop.

Why Docker made one sandbox for two places

Docker says it deliberately avoided building a separate cloud product with its own commands, because how much you can trust your agent shouldn't depend on where it happens to be running. As far as the company knows, no other agent sandbox works this way, it claims. Interactive work belongs on the laptop and work that takes hours belongs in the cloud, Docker argues, and most developers need both. For a limited time, new accounts receive 250 dollars in free Cloud Sandboxes credit.

What this changes for companies

For Taiwanese software teams that already use coding agents, Cloud Sandboxes turn agent runs into a metered cloud resource with the same isolation as on a developer's machine. The cost is easy to estimate, but two points deserve attention before adoption: code and data leave the company's machines for Docker's servers, and model usage is billed on top by the model provider. Teams should also set network policies and secrets per sandbox from the start, since those settings do not follow a sandbox when it moves.