Skip to content

OpenAI Codex

Draft · Version 0.3.0 · 8 pages · PDF · Word

You can give Codex standing instructions, skills, MCP servers, settings and permissions. Each one is a plain file, and the Codex CLI, the IDE extension and the ChatGPT desktop app all read the same files.

Most of these files live in your Codex home folder, ~/.codex unless CODEX_HOME says otherwise, or in your project. An Accidental Guide to: Agents, Harnesses, Skills, and Tools explains the basics.

Term Meaning
Instructions file Guidance Codex reads at the start of every session.
Skill Steps for one kind of task, read when the task comes up.
MCP server A program that gives Codex tools for another system.
Config file The file that holds settings such as the model and MCP servers.
Permissions What commands may change, and when Codex stops to ask you.
Field Text
Identify by AGENTS.md, in ~/.codex/ or anywhere in a project.
What Codex reads your AGENTS.md, then each one from the project’s root down to where you work. An Accidental Guide to: Agents, Harnesses, Skills, and Tools explains them.
So what Put guidance in the folder it’s about. The file closest to your work has the last word.
Field Short answer
Loads At the start of a session
Add with /init

Codex reads an instructions file every session. A skill’s steps are read only when the skill is used.

Field Text
Identify by A folder with a SKILL.md file.
What Codex reads the rest of a skill when your task matches its description, or when you type $ and its name. An Accidental Guide to: Agent Skills explains skills.
So what A project’s skills in .agents/skills/ reach everyone who works on it.
Field Short answer
Loads Description first
Check with /skills

A plugin packages skills, MCP servers and hooks to install together, using /plugins in the CLI.

Field Text
Identify by A [mcp_servers.docs] table in config.toml.
What Each server’s entry gives a command that starts it on your computer, or the url of a remote one. An Accidental Guide to: Model Context Protocol explains MCP servers.
So what Set a server up once and it works in the CLI, the IDE extension and the desktop app.
Field Short answer
Add with codex mcp add
Check with /mcp

The MCP server is the program. The tools it offers are what Codex calls.

Field Text
Identify by config.toml, in ~/.codex/ or in a project’s .codex/ folder.
What Settings are the choices that control how Codex runs, such as the model. Codex keeps them in config.toml, along with MCP servers and permissions.
So what Your own config.toml applies everywhere, and a project’s .codex/config.toml changes settings for that project.
Field Short answer
Loads When Codex starts
Override with -c for one run

Your config.toml sets defaults you can change. An administrator’s requirements.toml sets limits you can’t.

Field Text
Identify by sandbox_mode and approval_policy in config.toml.
What Codex has two permission settings. The sandbox decides what commands can touch: read-only, workspace-write or danger-full-access. The approval policy, on-request or never, decides when Codex stops to ask. An Accidental Guide to: Agents, Harnesses, Skills, and Tools explains permissions.
So what The sandbox keeps a wrong command away from everything outside your work. In a Git folder Codex starts with workspace-write and on-request, and elsewhere read-only.
Field Short answer
Network Off by default
Change with /permissions

Rules files in ~/.codex/rules/ say whether a command may run outside the sandbox: allow, prompt or forbidden.

Your own files apply in every project. A project’s files are kept in its repository.

Scope Thing Path
Your home folder Instructions ~/.codex/AGENTS.md, or AGENTS.override.md beside it
Your home folder Skills ~/.agents/skills/
Your home folder Config file ~/.codex/config.toml, which also holds MCP servers and permissions, with profiles beside it
Your home folder Hooks, rules, agents ~/.codex/hooks.json, ~/.codex/rules/, ~/.codex/agents/
This project Instructions AGENTS.md in any folder
This project Skills .agents/skills/ in any folder
This project Config file .codex/config.toml
This project Hooks, rules, agents .codex/hooks.json, .codex/rules/, .codex/agents/
Your organization Config file /etc/codex/config.toml, or defaults from the admin console
Your organization Skills /etc/codex/skills/
Your organization Requirements /etc/codex/requirements.toml, or the admin console
Codex Cloud Skills Only the repository’s skills. Personal skills aren’t synced.
Config file Where the value comes from, strongest first
1 A flag or -c on the command line
2 The project’s .codex/config.toml closest to where you work
3 A profile picked with –profile
4 Your ~/.codex/config.toml
5 Your organization’s defaults
6 Codex’s built-in defaults
Other setup When two places set it
Instructions Codex joins them, yours first, then the project’s from the root down, so the closest file has the last word.
Skills Codex offers all of them. Two with the same name both appear.
MCP servers Servers can be in your config.toml or a project’s, and a setting for the same server follows the numbered list.
Hooks Codex keeps all of them.
Rules The strictest match wins: forbidden, then prompt, then allow.
Codex Claude Code Claude Desktop Copilot CLI VS Code and Visual Studio
Instructions file CLAUDE.md Instructions AGENTS.md AGENTS.md
Skill Skills Skills Skills Skills
MCP server .mcp.json Connectors .mcp.json .mcp.json
Config file settings.json Settings settings.json Settings
Permissions settings.json Settings Settings Settings

Each of these tools has its own guide: An Accidental Guide to: Claude Code, An Accidental Guide to: Claude Desktop, An Accidental Guide to: GitHub Copilot CLI, and An Accidental Guide to: VS Code and Visual Studio.

These public sources were checked on 22 October 2026.

Source Publisher Address
Config basics OpenAI learn.chatgpt.com/docs/config-file/config-basic
Advanced Configuration OpenAI learn.chatgpt.com/docs/config-file/config-advanced
Custom instructions with AGENTS.md OpenAI learn.chatgpt.com/docs/agent-configuration/agents-md
Build skills OpenAI learn.chatgpt.com/docs/build-skills
Model Context Protocol OpenAI learn.chatgpt.com/docs/extend/mcp
Agent approvals and security OpenAI learn.chatgpt.com/docs/agent-approvals-security
Package your plugin OpenAI developers.openai.com/plugins/build/plugins
Hooks OpenAI learn.chatgpt.com/docs/hooks
Rules OpenAI learn.chatgpt.com/docs/agent-configuration/rules
Cloud environments OpenAI learn.chatgpt.com/docs/environments/cloud-environments
Managed configuration OpenAI learn.chatgpt.com/docs/enterprise/managed-configuration

A person writes and reviews every guide. AI does the grunt work.

Leave a star