Docker has extended its local sandbox infrastructure for coding agents to the cloud, addressing one of the most practical limitations of agentic development: the agent stops when you close your laptop.
The new Cloud Sandboxes let developers launch coding agents — Claude Code, Codex, Copilot, and others — on Docker-operated infrastructure using the same microVM-based isolation model as local sandboxes. Each agent runs in its own microVM with an independent kernel and Docker daemon.
The key feature is bidirectional migration: a sandbox can start locally, then be moved to the cloud when the developer steps away, and brought back later. Docker captures the filesystem and recreates the environment at the destination — it's a filesystem migration, not a running process handoff.
Cloud Sandboxes can be launched via CLI (`sbx --cloud run codex`) or web console, with pre-configured kits for popular agents. Security is handled through per-sandbox credentials and network policies, with a proxy that injects secrets without exposing them to the agent.
Pricing is per-second compute time. The standard tier (2 vCPUs, 4 GB RAM) costs $0.14/hour, with sizes ranging from $0.07 to $1.12/hour. Paused sandboxes incur no charges. Sandboxes default to one-hour lifetime, extendable up to 24 hours.
The move positions Docker as the runtime layer for agentic coding — not the agent itself, but the infrastructure that keeps agents running, isolated, and connected to the tools they need. Local sandboxes remain free and don't require Docker Desktop.
