Since October, Claude Code can change its own app. You describe what you want in a sentence, such as "show the current client and branch above the prompt" or "ask me before any command touches the production server", and Claude writes a small piece of code that does it. Anthropic calls these mods.
It is a lot of flexibility, and it needs care: a mod is code that runs with your permissions, and nothing walls it off from the rest of your computer. This guide covers what a mod can do, six that a small business would actually use, the precautions to take, and what OpenAI offers instead in Codex.
What a mod is, in plain words
A mod is a small program, written in JavaScript or TypeScript, that lives inside a plugin and reacts to what happens in Claude Code: a message being sent, a command about to run, the screen being drawn. Claude writes it for you, but it helps to know what it is, and someone who reads code should look at it before you rely on it.
It is a different thing from what we covered in our desktop app guide. Connectors and skills give the agent new abilities from the outside: access to your shop, a set of instructions for a recurring task. A mod changes the app itself, what you see and what is allowed to happen. You may already be using a few: Claude Code ships some built in, such as the one that draws charts in the conversation.
Mods work in Claude Code in the terminal, in the Code tab of the Claude desktop app and in the VS Code extension. They do not apply to the Chat or Cowork tabs, and what they can display varies with the version and the type of session.
What a mod can change
The list is longer than most people expect. In the desktop app, a mod can:
- Add a panel beside the conversation, with tabs, buttons and text fields, that updates while you work.
- Add a strip above the prompt box for something you want in view at all times.
- Hold a command before it runs, ask you a question, or refuse it with an explanation.
- Add your own slash command that runs straight away, without waiting for Claude, even while Claude is busy.
- Redraw parts of the existing interface, such as messages, tool calls and the waiting indicator.
- Change the colors of the Code tab and the backdrop behind the conversation.
How you make one: you ask
You do not need a developer setup. In a Claude Code session on your own computer, describe the mod: "make a mod that shows the current git branch above the prompt" is Anthropic's own example. Claude uses a built-in skill for writing mods, and there is no build step or extra software to install.
In the default permission mode, Claude Code asks you to approve each file Claude wants to create, and after the first one asks whether to turn on hot reloading for the session. Auto mode can skip both questions. Say yes and the mod loads when Claude's turn ends. If it is not quite right, say what to change: the mod reloads after each turn that edits it, so you see the new version straight away.
Keep the ones you like
This is the detail most people miss. A mod Claude writes lives in a folder that belongs to that session, loads only in that session, and is deleted with the session's other files once the session is older than your cleanup setting. It comes back if you resume that session, but not in a new one.
To keep a mod, ask Claude to copy it out of the session folder and turn it into a plugin you install, which then loads in every session. This is the one step where a non-developer may want help the first time. Once it is done, the mod behaves like any other plugin: it has a version number, and you update or remove it from the plugin settings.
Six mods worth asking for
These are ideas, not products: each one is a sentence you can give Claude. The first two come from Anthropic's own examples, the others are what we would build for a small team.
- A production guardrail: hold any command that mentions your live server or database, show what it would change, and wait for you to press Proceed or Cancel. It can miss a command written differently, so it does not replace limiting what the agent can access.
- A deploy panel: your site's latest deployments beside the conversation, updating as they pass or fail.
- A "today" panel: open orders, bookings or support tickets, read from a file or a tool your folder already uses, refreshed every few minutes.
- A client strip: which client, project and branch this session is working on, above the prompt, so nobody edits the wrong site.
- A /closing command: start your end-of-day routine, such as a summary of what changed and what is still open, with one word. If it asks a model for the summary, that takes a moment and uses your plan.
- Your brand colors on the Code tab, in light and dark mode. Little value, but a low-stakes first mod to try.
Share them with your team
Because a mod is a plugin, it travels the way plugins do, and Anthropic documents a route for each size of team.
- Two or three people: send the plugin's folder or a zip of it.
- A team: list it in your own marketplace, which can be a private repository with one folder per plugin.
- A whole organization: an administrator installs it for everyone through managed settings.
Read this before installing any mod
Anthropic is direct about it: mods are not sandboxed. A mod runs as you. It can read and write your files, start programs, make network requests and read environment variables and settings, including any API key kept there. It can see and rewrite every prompt and command, approve some commands before you are asked, and spend your usage by calling a model.
What it cannot change is the permission prompt: a mod can restyle most of the interface, but not what a permission request shows you. On Team and Enterprise plans, a built-in guard also keeps your deny rules and required confirmations in force when a mod approves a command, but a mod's own access to files and programs is still not walled off. Treat a mod like any software you install, and follow a few rules.
- Have every mod reviewed by someone who reads code before you rely on it, including one Claude wrote for you, and install others only from authors you trust.
- Ask for a mod in the default permission mode, so you approve each file. In auto mode, Claude Code can load a new mod without asking.
- When reviewing a mod, run claude plugin validate: it lists the events a mod reacts to and what it asks Claude Code to do, and catches structural errors, but it does not certify that a mod is safe.
- Know the off switches: disable a mod in the plugin settings, or start a terminal session with --safe-mode to turn off installed mods for that session. Built-in mods stay on.
- A guardrail mod is a convenience, not a security boundary. Real protection is in the live system itself: accounts with only the permissions they need, and production credentials kept out of the account Claude Code runs under.
Does Codex have an equivalent?
Not as one feature. OpenAI covers some of the same ground in Codex and the ChatGPT desktop app, split across several tools, and its documentation does not describe asking the app to change its own interface.
- Themes: a built-in setting for base theme, accent, background, text colors and fonts, which you can share. Simpler than a mod for that one job.
- Hooks: scripts or connected tools that run at points such as before a tool is used, close to the guardrail idea. Hooks you add yourself must be reviewed and trusted before they run. They can show a warning message, but not the panels and buttons a mod can draw.
- Panels beside a conversation: possible through plugin extensions, which need an MCP server and a custom interface, so technical work. ChatGPT can build workflow plugins from a conversation without code, but those do not change the app's interface.
- Reusable routines: skills, which are instructions for the model rather than code that runs instantly.
When a mod is the wrong tool
Because a mod is code that runs as you, it should be the last tool you reach for. Anthropic's own documentation points to simpler tools for most needs.
- You keep pasting the same instructions: write a skill, which is a plain text file.
- Claude needs data from another system: add a connector, or a command-line tool it can run.
- You always want to block one specific command: add a permission rule, which takes no code.
- Your work depends on it: the events and methods mods use can change between releases, so do not hang a critical process on one.
Where to start this week
Pick one mod from the list, ask Claude for it in a session with the default permissions, have someone who reads code look at it, and use it for a week. If it earns its place, have it turned into a plugin. If you followed our folder guide, the client strip and the today panel are good first choices, because they show information your folder already holds.
Where we help
We set up Claude Code and Codex for businesses that want AI working on their website and internal tools: the folder, the permissions, the scheduled tasks, and now the few mods that make the app fit how the team works. We also review mods before they reach anyone's computer, which is the part we would not skip.
