
MCP in Codex is how OpenAI's Codex programming agent calls tools and reads data from external MCP servers, declared via the codex mcp command or the [mcp_servers.<name>] table in the config.toml file. Besides MCP, Codex has three other extension mechanisms: skill (instruction packages that can include scripts), plugin (pre-packaged bundles of many things, installed from the plugin marketplace) and hooks (commands that run at specific points in the session lifecycle). This post is based on Codex documentation on learn.chatgpt.com and the --help output of Codex CLI version 0.162.1, read on 03/11/2026; features may change across versions. This post makes no claims about the safety or effectiveness of any utilities and does not provide sample hook configurations because we have not yet written any blocking hooks. The post covers how to add MCP servers, skills, plugins, hook points, a comparison table, risks, a decision tree and real experience from OPCAIOS. If you're not yet clear on MCP, first read What is MCP.

Four mechanisms to extend Codex and where to declare each. Diagram: OPCAIOS.
How do I add an MCP server to Codex using the codex mcp command?
Adding an MCP server to Codex via the command line is the quickest way: the codex mcp add command writes the declaration to ~/.codex/config.toml instead of editing by hand. According to codex mcp --help from Codex CLI 0.162.1 (run on 03/11/2026), the command group includes list, get, add, remove, login and logout. The add command takes a name, then either --url for a streamable HTTP server, or -- followed by a command to start a local server (stdio). The --env KEY=VALUE flag is only for stdio servers, while --bearer-token-env-var is only for HTTP servers and takes the name of an environment variable holding the token, not the token itself. Example from Codex MCP documentation:
codex mcp add context7 -- npx -y @upstash/context7-mcp
codex mcp list
codex mcp login <server-name>
In Codex's terminal interface, type /mcp to see which servers are running.
What keys does an MCP server declaration in config.toml contain?
Declaration in config.toml gives finer control than the add command. According to Codex's MCP page (learn.chatgpt.com/docs/extend/mcp, read on 03/11/2026), each server is a [mcp_servers.<name>] table. A stdio server needs command, and can have args, env, cwd. An HTTP server needs url, and can have bearer_token_env_var and http_headers. Notable common keys: startup_timeout_sec (default 10 seconds), tool_timeout_sec (default 60 seconds), enabled, required, enabled_tools and disabled_tools.
[mcp_servers.context7]
command = "npx"
args = ["-y", "@upstash/context7-mcp"]
tool_timeout_sec = 60
[mcp_servers.vi-du-http]
url = "https://mcp.example.com/mcp"
bearer_token_env_var = "VI_DU_TOKEN"
enabled_tools = ["search"]
The second server is a hypothetical example. The documentation notes that OAuth client secrets, if declared, will appear in lowercase in the file, so keep the file private and outside of git.
What is a skill in Codex and where do I put it?
A skill in Codex is a directory containing a SKILL.md file (instruction plus metadata), optionally with scripts/, references/ and assets/, helping Codex carry out a repeatable process correctly. According to the build-skills page in Codex documentation (read on 03/11/2026), the beginning of SKILL.md must have name and description, and skills follow the open agent skills standard. Codex looks for skills in four layers:
- Repository:
.agents/skillsin the current directory, parent directory and repository root. - User:
$HOME/.agents/skills. - Administrator:
/etc/codex/skills. - System: bundled with Codex, including
skill-creatorandskill-installer.
Call a skill with /skills or type $ followed by the name; Codex also auto-selects a skill when it matches description. Codex only loads the name and description first, reading the full SKILL.md when that skill is chosen; the initial list by default takes up at most 2% of the context window. So description should be clear about when to use it and when not to. The general concept of skills appears in the post What is skill and how to use skill.
How does Codex Plugin and its plugin marketplace work?
A Codex plugin is a pre-packaged set of capabilities to install once: according to the Plugins page in the Codex documentation (read on 03/11/2026), a plugin can include a skill, MCP server, browser utility, and hooks. Codex CLI and desktop application can use plugins, but IDE utilities cannot. There are three ways to install and manage:
- In Codex CLI, type
/plugins, install from a configured marketplace, then open a new session. - Use the command:
codex plugin add <plugin>@<marketplace>to install,codex plugin listto view,codex plugin removeto uninstall (percodex plugin --help). - Add a custom marketplace:
codex plugin marketplace addaccepts a local path,owner/repo, or Git URL; hasupgradeto refresh andremoveto delete.
The Plugins directory has an OpenAI tab, a workspace tab, and a personal tab. To compare with commands, skills, plugins, and MCP on Claude Code, see the post slash command, skill, plugin or MCP; this post does not repeat that section.
At what points do hooks in Codex run?
Hooks in Codex are commands you declare that run automatically at fixed points in a session, unlike guidance in AGENTS.md which is only a suggestion to the model. According to the Hooks page (learn.chatgpt.com/docs/hooks, read on 03/11/2026), Codex has 14 points: SessionStart, SessionEnd, UserPromptSubmit, PreToolUse, PostToolUse, PermissionRequest, PreCompact, PostCompact, SubagentStart, SubagentStop, Stop, Interrupt, PreModelRequest, and PostModelRequest. Hooks are placed in ~/.codex/hooks.json, in the [hooks] table of config.toml, or in .codex/ of the repository; plugins also carry hooks along. Three important points:
- Only some points can block operations:
PreToolUsecan reject a command,UserPromptSubmitblocks a prompt,PermissionRequestallows or denies; exit code 2 with a reason is one way to block. PostToolUseblocks by replacing the result only, it does not undo the command already run; background hooks and points likeSessionEnd,Interruptcannot block.- Hooks not created by an admin must be reviewed and trusted before running; modifying a hook requires trusting it again. Use
/hooksto view, trust, or disable.
Hooks are on by default, disable with hooks = false in [features]. For hooks on Claude Code, see hooks in Claude Code.
Compare MCP, skill, plugin, and hooks in Codex
Four mechanisms solve four different problems. Summary table according to documentation read on 03/11/2026:
| Mechanism | Solves what problem | Declare where | Who decides when to run |
|---|---|---|---|
| MCP server | Let Codex call tools, read external system data | codex mcp add, [mcp_servers.<name>] |
Model chooses to call tools, by approval mode |
| Skill | Teach Codex a repeatable process | SKILL.md in .agents/skills |
You call with $, or model chooses by description |
| Plugin | Install a package of skill, MCP, hooks at once | /plugins, codex plugin add |
Depends on components inside |
| Hook | Run a command surely at a point, can block | hooks.json, [hooks] |
System, each time the point is reached |
Way to remember: MCP is an assistant's hand, skill is a notebook, plugin is a packed box, hook is a latch.
What are the risks when installing strange utilities for Codex?
The main risk when installing strange MCP server, skill, or plugin is giving code from others the power to run on your machine and read the data Codex can read. Five specific risks:
- MCP server stdio is a program that runs on your machine; commands like
npx -ydownload and run the latest package without asking again. - Tool descriptions and responses from MCP server go straight into model context, so they can contain prompt injection instructions.
- Skill can come with scripts; Codex runs that script in a sandbox and current approval, but if you already gave broader permissions then the script has those permissions.
- Plugin packs many things at once, including hooks that must be reviewed and trusted before running.
- Token and client secret declared in
config.tomlleak if the file is committed.
Measures: install from sources where you can read the code, lock package versions, use enabled_tools to turn on only needed tools, keep tokens in environment variables, maintain strict approval when trying. See also security when using AI to write code.
Decision Tree: Which Mechanism Should I Use?
Ask in sequence, stop at the first question answered "yes":
- Must this happen for certain, regardless of which decision model (blocking dangerous commands, logging)? Use a hook.
- Does Codex need to read or write to an external system (GitHub, database, calendar)? Use an MCP server.
- Is this a multi-step process you often explain again? Write a skill.
- Do you need to carry skills, MCP, and hooks to another machine or share with a team? Package it as a plugin, or install a ready-made plugin.
If it's just a general rule like "reply in Vietnamese", put it in AGENTS.md and you're done, no mechanism needed.
OPCAIOS Experience
OPCAIOS uses Codex CLI version 0.162.x and the Codex desktop application. These things are real, checked on 03/11/2026:
- All enabled plugins come from OpenAI's plugin marketplace: browser, visualization, Gmail, Google Calendar, Slack, documents, PDF, spreadsheets, and presentations. The machine also has Codex's Computer Use application installed. We haven't tracked the frequency of using each plugin yet, so we just say "enabled".
- Codex hooks are configured at session start, send prompt, before and after tools, permission request, subagent, and stop, but only to report session status to a terminal management tool. No hooks block commands yet.
- A plugin code review folder exists but is empty, so we can't claim we use Codex for code review regularly.
What we haven't done: haven't measured which plugins save time, haven't put any protective hooks in place of manual discipline.
Frequently Asked Questions
If I add an MCP server to Codex CLI, can I use it in the desktop application?
Yes, according to Codex MCP documentation read on 03/11/2026: the desktop application, Codex CLI, and IDE utilities all share the MCP configuration in ~/.codex/config.toml, so adding it once with codex mcp add makes it available to all three. The web version of ChatGPT doesn't read local config files; accessing MCP tools from the web goes through plugins. Configuration in .codex/config.toml in a project only takes effect when the project is trusted.
Can I reuse Claude Code skills in Codex?
The format is similar: both use a directory with SKILL.md containing name and description, and Codex documentation says skills are based on the open agent skills standard. The difference is location: Codex looks in .agents/skills in the repository and $HOME/.agents/skills, not Claude Code's own directory. Skills that call tools specific to one agent need adjustment; we haven't verified sharing them.
Can Codex hooks block dangerous commands?
Possibly, at the right point. According to the Hooks page read on 03/11/2026, the PreToolUse hook can reject a command before it runs, and the PermissionRequest hook can reject a permission request; exit code 2 with a reason is one way to block. The PostToolUse hook can't undo a command that's already run. Background hooks can't block anything, and all hooks must be trusted before they run. OPCAIOS hasn't written blocking hooks so we don't provide an example.
Should I install many plugins at once?
No, don't install everything you see. Each plugin may carry MCP servers, skills, and hooks—meaning it adds tools for the model to choose, adds descriptions that take up context, and adds code that runs on your machine. Enable a plugin when you have a specific task, read what it includes, check hooks in /hooks, remove it with codex plugin remove when you're done. The documentation notes that uninstalling a plugin in ChatGPT doesn't automatically close the plugin's MCP connection.
Where should someone who doesn't know programming start extending Codex?
Start with AGENTS.md for general rules, then install one official plugin from the OpenAI tab in the Plugins folder for what you actually need, instead of writing your own MCP server or hook. Once familiar, try $skill-creator to let Codex compose a skill from your description. Save hooks for last since they need shell command knowledge. The article safe process for people who don't know code has the fundamentals. For direct guidance, see OPCAIOS's AI application training course.
Source: Codex documentation on learn.chatgpt.com (pages extend/mcp, build-skills, plugins, hooks; redirected from developers.openai.com/codex) and codex --help, codex mcp --help, codex plugin --help from Codex CLI 0.162.1, read on 03/11/2026. Not verified: detailed JSON structure of hooks.json, sharing skills between Codex and Claude Code, plugin behavior in IDE utilities in later versions; features may change across versions. Experience is OPCAIOS's.
Translated from Vietnamese with AI. Read the original.
Want to apply this in your business?
Book a free 30-minute call (Vietnam time, in English or Vietnamese).
