OpenAI Codex
Codex
Leave the task to a coding agent running in the real codebase: it understands the repository, changes files, runs commands, and delivers reviewable diffs. This page explains the most important working patterns for getting started with Codex.
Cloud
Ideal for parallel tasks and PR-level delivery
CLI
Suitable for local terminal and real verification
IDE
Suitable for collaboration while looking at code
Workflow
Codex
Multitasking in parallel
Separate tasks that are independent of each other and let multiple Codex threads process them in parallel.
Security verification
Each finish requires command output and reproducible steps, especially login, payment, and data writing.
Rule capture
Place project commands, design exclusion zones, and release criteria into AGENTS.md or skills.
Prompt pattern
To reproduce this bug, first explain the possible reasons; change only the minimum necessary files; run relevant tests; and finally give me the reproduction steps, summary of changes, and remaining risks.
How to use it
Understand the tool, then read the docs
This entry page keeps the tool overview, use cases, and first steps up front, then organizes the documentation index for deeper reading.
Cloud tasks
Give Codex a clear target, wait for it to read the code, modify the files, run checks, and return auditable results.
Local CLI
Start in the repository root and directly let Codex read the project, execute the command, and retain the complete work log.
Code review
Treat output as a PR candidate: focus on diff, command output, permission boundaries, and rollback methods.
Skills
Turn repetitive processes into skills and let Codex execute according to the same set of engineering rules every time.
Install and log in
Get started with the Codex App, CLI or IDE extension. CLI is suitable for local projects, and App is suitable for managing multiple tasks in parallel.
Reduce tasks into one project card
Write down the goals, context, constraints, acceptance criteria, and commands that must be run. Don’t start with “Help me optimize it.”
Let it read first, then let it change
The first response simply asked Codex to read the repository and list a plan. After confirming the orientation, allow it to apply changes.
End with evidence
Builds, tests, browser checks, screenshots and diffs are all evidence. Don't merge "done" without evidence.
Codex documentation
Documentation index
Migrate and organize the information architecture of the OpenAI Codex documentation: getting started, concepts, usage, configuration, automation, and best practices.
Getting Started
Codex Overview
Learn how Codex acts as a software engineering agent, completing tasks, modifying files, and delivering auditable results at every entry point you write.
ReadGetting Started
Codex quick start
Start Codex with a minimal closed loop: install, log in, open project, describe task, review diff, run verification.
ReadGetting Started
Codex use cases
Think of Codex as a teammate that can handle engineering tasks in parallel: let it read code, fix bugs, add tests, migrate pages, and prepare PRs.
ReadGetting Started
Overview
Learn how Codex acts as a software engineering agent, completing tasks, modifying files, and delivering auditable results at every entry point you write.
quick start
Start Codex with a minimal closed loop: install, log in, open project, describe task, review diff, run verification.
use case
Think of Codex as a teammate that can handle engineering tasks in parallel: let it read code, fix bugs, add tests, migrate pages, and prepare PRs.
Import into Codex
The key to letting Codex understand the project is not to stuff a lot of text at once, but to provide repositories, rules, task goals and verification methods.
concept
prompt
High-quality prompts should let Codex know the goals, boundaries, acceptance criteria, and closing format.
Customize
Capture recurring preferences into rules, configurations, and skills to keep the Codex consistent across sessions.
memory
Memory is used to retain stable preferences, but project facts should still be based on repository files, documentation, and command output.
Sandbox and Approval
Understand how the Codex sandbox restricts file, command, and network access, and how it works with approval policies to control the degree of autonomy of local agents.
subagent
Subagents are suitable for breaking large goals into parallel small tasks, but each subtask should have clear boundaries and unified acceptance criteria.
Workflow
A stable workflow usually includes: reading code, making plans, executing small steps, running verification, reviewing differences, and summarizing risks.
Use Codex
Application
Codex applications are suitable for managing multiple engineering tasks, viewing agent execution records, and converting results into reviewable development output.
IDE extensions
IDE extensions are great for collaborating alongside the code you're looking at, reducing context switching.
command line
CLI is suitable for local repositories, real commands and reproducible verification.
Web version
The web version is suitable for submitting remote tasks, viewing the progress of long tasks and managing multiple sessions.
GitHub
When connecting Codex to GitHub, the goal is to shorten the distance between issues and PRs while retaining human review.
Slack
Slack is great for teams to initiate tasks and receive status, but the final merge should still lead back to the code review process.
Linear
Linear is suitable for converting product requirements into engineering tasks, and then Codex handles verifiable subtasks.
Configuration
Configuration file
Configuration lets Codex know the project's default commands, permission preferences, and working methods.
Permissions
Learn how to migrate Codex permission profiles for file system, network, path, glob, Unix socket and old sandbox settings.
AGENTS.md
AGENTS.md is a project-level rule entry, suitable for recording commands, architecture, coding style and collaboration boundaries.
MCP
MCP can connect internal tools, documents, work orders, and data sources to Codex, but it must first deal with permissions and data boundaries.
Skills
Skills are used to encapsulate repetitive workflows so that Codex executes them in stable steps every time.