Don't Panic — OpenAI Codex CLI

OpenAI Codex CLI is a coding agent in a terminal. Instead of taking a code snippet away to a chat window and returning with a hopeful suggestion, it can inspect the repository in front of it, edit files, and run the tools already used by that repository. This is useful. It also means the small glowing terminal has been promoted from conversation partner to colleague with access to the stationery cupboard.

The important shape is a loop: describe the task, let Codex inspect the relevant files, review its proposed commands and diff, then verify the result. The model can carry context across the session, but it does not acquire the missing requirements in an ambiguous request. A precise task and a small scope make a change easier to assess; a test run makes the result more than an eloquent theory.

Sandbox modes set where the agent can operate. Read-only is for inspection. Workspace-write is for work inside the project. Full-access reaches more of the machine and therefore deserves particular caution. The companion control is the approval policy, which decides when Codex asks before an action. The practical rule is pleasantly unromantic: begin with the narrowest access that permits the task, then review commands before expanding it.

Codex also needs the local rules that humans usually learn by overhearing conversations. AGENTS.md supplies project instructions such as architecture, conventions, and test commands. It tells the agent which shared client is the shared client, rather than asking it to locate enlightenment somewhere in a directory tree. MCP connections add external tools, but each connection expands what a session can reach, so they belong in the same review habit as shell commands.

Start with the Intro for the mental model and the Slides for the relationships between prompts, permissions, instructions, and tools. Use the Cheatsheet when choosing flags or configuration. Then take the Exercise: it turns the loop into a small, observable change in a disposable repository. The durable method remains prompt, review, approve, iterate. It is not magic, but it is much more useful than magic when the tests fail.