Start here

Telling it what you want

Running it without sitting there

Making it yours

More than one Codex

Trust

Making it real

Codex is an agent that edits your code and runs your commands inside a sandbox you control. This course starts with that sandbox, proves what it blocks by running it, then builds up: approvals, AGENTS.md, rules, non-interactive runs, reviews, skills, plugins, MCP, hooks, subagents and the cloud.

OpenAI Codex

The one runnable lesson

OpenAI Codex course: a complete, practical tutorial

What you are going to build

You can follow along either way

Installation: install the OpenAI Codex CLI

Installation: what you get

One turn

It covers what it runs

How Codex works: the loop and the four surfaces

How Codex works

And neither is anything outside the folder

What is actually enforced

The Codex sandbox: what it blocks, shown by running it

Sandbox: shown by running it

The habit to keep

Codex permission profiles: read-only, workspace, full access

Permission profiles

Being asked is not being protected

Codex approval policy: when the agent asks permission

Approval policy: when it stops and asks

It is a strong suggestion

AGENTS.md in Codex: instructions it reads every session

AGENTS.md: standing instructions

How wide to make a pattern

Codex rules: control which commands can run outside the sandbox

Rules: what may leave the sandbox

When to stop retyping it

Prompting Codex: how to give OpenAI Codex a good task

Prompting Codex

Unattended means explicit

codex exec: run OpenAI Codex non-interactively

codex exec: one prompt, no session

Sessions and directories

Codex sessions: resume, fork and queue work

Sessions: resume, fork and queue

Ask for the review you want

Run a code review with codex review

codex review: a review on demand

They do not travel

Worktrees and environments in OpenAI Codex

Worktrees and environments

Commit the limits

The Codex config.toml file and how settings are resolved

config.toml

Why the description has to earn its place

Do it twice first

Codex skills: package a workflow it can repeat

Skills

It is code you are running

Codex plugins: install and share skills and connectors

Plugins

Tool lists are not free

Add MCP servers to Codex

MCP servers

Where the real boundary is

Codex hooks: block a tool call with your own program

Hooks: blocking a tool call

Parallel is not free

Codex subagents: parallel work in one session

Subagents

Internet access is a choice

Codex cloud: run tasks in isolated environments

Codex cloud

Tickets are untrusted input

GitHub, Slack and Linear integrations for OpenAI Codex

GitHub, Slack and Linear integrations

Before you widen anything

Security model of OpenAI Codex: sandbox, approvals, network

Security model

Fatigue is the real risk

Codex auto-review: a reviewer agent at the sandbox boundary

Auto-review

Nothing reviews a memory

Codex memories versus AGENTS.md

Memories: what it remembers

Access is per account

Models and fast mode in OpenAI Codex: speed and cost

Models and fast mode: speed and cost

Build it over a fortnight, not an afternoon

OpenAI Codex: a complete repository setup, end to end

Repository setup: end to end

Telling it what you want quiz

With on-request approvals, when does the agent stop and ask?