Codex CLI vs Codex Cloud: Picking The Right Surface

Codex CLI runs on your machine, sees your files, your env vars, and your installed tools. Codex Cloud runs in an OpenAI sandbox — a clean container with whatever you installed in its setup script. The CLI is for tight loops; the cloud is for delegation.

When the CLI wins

• You need access to a local database, VPN, or staging environment
• You want sub-second loops between prompt and result
• You are pair-programming and need to interrupt the agent often
• Your repo has uncommitted state you want the agent to see
• You need access to local secrets that should never leave your laptop

When the cloud wins

• You want to start a task and walk away
• You want three or four PRs in flight at once
• The work is well-bounded enough to run with no human in the loop
• You want a clean repo state with no laptop-specific oddities
• You want the result to land as a GitHub PR with full diff review

Applied exercise

1. 1List the last five coding tasks you delegated to an AI
2. 2Mark each one C (CLI-fit) or X (cloud-fit) using the criteria above
3. 3If your list is mostly C, your delegation is too small. If it is mostly X, you are missing tight-loop work
4. 4Aim for a healthy mix — both surfaces sharpen each other

The big idea: pick the surface that matches the task's shape. Tight loop and dirty state — CLI. Long horizon and clean PR — cloud.