CDX-201b · Module 1

The Cloud Execution Model

4 min read

Codex Cloud is fundamentally different from local Codex CLI. When you run a task locally, Codex operates on your filesystem, in your shell, with your resources. When you delegate to Codex Cloud, the task runs in a remote sandbox hosted by OpenAI — an isolated environment with its own filesystem, compute, and network policies. You submit a task, the cloud executes it, and you review the results asynchronously. The interaction model shifts from real-time collaboration to async delegation.

Each cloud task gets its own microVM — a lightweight virtual machine that boots in seconds, runs the task, and shuts down when complete. The microVM clones your repository, applies your AGENTS.md configuration, installs dependencies, and executes the task in an environment that mirrors your local setup as closely as possible. When the task finishes, the results (diffs, test output, logs, screenshots) are available for review through the Codex web interface, CLI, or mobile app.

# Delegate a task to Codex Cloud
codex cloud "refactor the authentication module to use JWT refresh tokens"

# One-shot cloud execution
codex cloud exec "run the full test suite and report failures"

# Check cloud task status
codex cloud status

# Review a completed task
codex cloud diff <task-id>
codex cloud logs <task-id>

Do This

  • Use cloud execution for tasks that take more than a few minutes
  • Delegate independent tasks in parallel — each gets its own microVM
  • Review cloud results carefully before merging — async means less oversight during execution

Avoid This

  • Use cloud for quick interactive tasks — local Codex is faster for rapid iteration
  • Assume cloud mirrors your local environment exactly — dependency versions may differ
  • Skip reviewing cloud output because "it ran in a sandbox so it must be safe"