> ## Documentation Index
> Fetch the complete documentation index at: https://docs.zerodrift.com/llms.txt
> Use this file to discover all available pages before exploring further.

# ChatGPT and Codex Plugin

> Install the ZeroDrift plugin once from the ChatGPT plugin marketplace and use it in ChatGPT and Codex. Install steps, how it works, what each distribution enforces, and troubleshooting.

The ZeroDrift plugin for ChatGPT and Codex checks drafts the assistant writes against your company policies. ZeroDrift returns findings and never rewrites your text.

**Install it from the ChatGPT plugin marketplace.** This is the best way to use the plugin, and the simplest way to install it for most organizations. Installing it from the ChatGPT web plugin page installs it on all your devices for that account, and makes it available everywhere, in ChatGPT and in Codex.

[![](https://docs.zerodrift.com/images/openai-plugin/openai-logo-white.png)
Add to ChatGPT](https://chatgpt.com/plugins/plugin_asdk_app_6af4ea9184388191b9ee1ff948e85d34)

Opens the ZeroDrift listing in the ChatGPT plugin directory. Installing it also enables the plugin in Codex.

After you install, try a [task guide](#next-steps).

## Install the plugin

The plugin is available to everyone, in all regions, on any OpenAI plan. These steps are the only install you need.

**1. Find the plugin**

Open the [ZeroDrift listing](https://chatgpt.com/plugins/plugin_asdk_app_6af4ea9184388191b9ee1ff948e85d34), or open **Plugins** and search for **ZeroDrift**.

![The ZeroDrift plugin listing in ChatGPT, with the Install plugin button, example prompts, and the Apps and Skills sections](https://mintcdn.com/zerodrift/WR3oWCMjKWZnWdS7/images/openai-plugin/chatgpt-plugin-listing.png?fit=max\&auto=format\&n=WR3oWCMjKWZnWdS7\&q=85\&s=93a4f36712415a71b6350fae046457ed)

**2. Install**

Select **Install plugin**, at the top right of the listing.

**3. Connect your account**

ChatGPT sends you to ZeroDrift's **Connect ChatGPT** page, which shows the app (ChatGPT), where it returns to (chatgpt.com) and your workspace. Select **Approve** to let ChatGPT check messages with your ZeroDrift policies, or **Deny** to cancel.

![ZeroDrift Connect ChatGPT approval page showing the app, the return address, the workspace, and Deny and Approve buttons](https://mintcdn.com/zerodrift/WR3oWCMjKWZnWdS7/images/openai-plugin/chatgpt-connect-approve.png?fit=max\&auto=format\&n=WR3oWCMjKWZnWdS7\&q=85\&s=7c1dbcf6c4be81a2a466420c4b7743db)

**4. Confirm**

ChatGPT shows "ZeroDrift plugin installed" with **Try now**, **Settings** and **Try in chat**. Ask for the ZeroDrift status to confirm the connection.

![The ZeroDrift plugin page with a "ZeroDrift plugin installed" confirmation and the Settings and Try in chat buttons](https://mintcdn.com/zerodrift/WR3oWCMjKWZnWdS7/images/openai-plugin/chatgpt-plugin-installed.png?fit=max\&auto=format\&n=WR3oWCMjKWZnWdS7\&q=85\&s=d0d5a0e8f92884b777c790b022da9030)

### Codex: nothing more to install

You do not install the plugin again for Codex, and you usually do not need to do anything there. If Codex does not show the plugin or asks you to sign in, use this as a fallback: sign in to the ZeroDrift connection from a terminal, then approve access in the browser.

```bash theme={null}
codex mcp login zerodrift
```

Ask for the ZeroDrift status to confirm the connection. You also need this command, and the hooks version, only if you want hooks. See [hooks version](#hooks-version).

### Install for your workspace (admins)

Only organization admins of the OpenAI account can see this screen. From here you can install the plugin for the whole organization, or make it available to the whole organization.

**1. Open the plugin page**

Open **Plugins > ZeroDrift**, select the **...** menu, and choose **Manage for workspace**. The menu also has **Copy link** and **Uninstall**.

![The ZeroDrift plugin page with the menu open, showing Copy link, Uninstall and Manage for workspace](https://mintcdn.com/zerodrift/WR3oWCMjKWZnWdS7/images/openai-plugin/plugin-menu-manage-for-workspace.png?fit=max\&auto=format\&n=WR3oWCMjKWZnWdS7\&q=85\&s=247decb396f9c3d526fad18240e2f68c)

**2. Set availability**

In **Workspace Settings > Plugins > ZeroDrift**, the plugin shows **Enabled**. Under **Availability**, the **Workspace default** applies to everyone in the workspace. Choose **Available** or **Installed**.

![Workspace Settings for the ZeroDrift plugin, showing Availability, App, Sites, MCP Servers, Skills and Information](https://mintcdn.com/zerodrift/WR3oWCMjKWZnWdS7/images/openai-plugin/admin-workspace-settings.png?fit=max\&auto=format\&n=WR3oWCMjKWZnWdS7\&q=85\&s=89b0f540a3ed212cb0f2399ab348b6b1)

**3. Review the rest**

The page also lists the app URL, **Permissions** (workspace default), **Tools** (3), **New tools** (**Enable all new tools**), and a **Sites** toggle, "Allow site visitors to use this plugin". It shows the MCP server `zerodrift`, the `status` and `validate` skills, and the listing information (Catalog **Public**, Version 0.1.9).

Install the plugin from the directory. Adding only the MCP endpoint does not include the plugin's skills, so the pre-send check instructions would be missing.

The plugin connects to the remote MCP service at `https://mcp.zerodrift.ai/mcp` using OAuth. You do not copy an API key.

## Before you start

* An existing ZeroDrift account, and authorization to its workspace. Checks use your workspace's live policies.
* You can sign in to ZeroDrift Command in a browser and approve access.

The plugin works on any OpenAI plan. In ZeroDrift, workspace Admins, Owners and Developers can install the plugin, grant OAuth consent and access the workspace.

> [!NOTE]
>
> In ChatGPT Business and Enterprise workspaces, owners and admins control plugin installation under **Admin console > Plugins**. The policy options are **Available for installation**, **Installed for workspace**, **Customize by role** and **Disable**. Member permissions such as **Use plugins** and **Upload plugins with custom MCP servers** are set there too ([workspace plugin controls](https://learn.chatgpt.com/docs/enterprise/apps-and-connectors)). Workspaces in FedRAMP show **Apps** instead of **Plugins**.

## How it works

ZeroDrift does not connect to your mailbox or to the sites where you publish. Your assistant already does that. ZeroDrift can validate any content, or plugin reference content, that Codex or ChatGPT has access to. If ChatGPT or Codex can read or write it, the assistant can hand it to ZeroDrift. ChatGPT formats the text itself so that it passes the validation API.

1. You ask the assistant for a task in plain language, such as "Check this email with ZeroDrift".
2. The assistant gets the content, either from your message or through another connector it has, such as Gmail.
3. The assistant sends the text to a ZeroDrift tool (`validate`, `check_send` or `status`).
4. ZeroDrift checks the text against your active rules and returns the result.
5. The assistant shows you the result and each flagged line, with the rule and policy it breaks.

The assistant is the interface. ZeroDrift is the validation engine behind it. The plugin's `validate` skill checks a draft and its `status` skill confirms the connection. The server also has a `check_send` tool, which the hooks version's hooks call before a Gmail action.

The assistant shows the outcome and next step, then each flagged line with its severity, rule and policy. It suggests wording only if you ask, and ZeroDrift has not checked that wording. If a result is Flagged, the validate skill tells the assistant not to send, post or publish the content and to let you decide.

Results also appear in the Activity feed in ZeroDrift Command. Each check is listed with its status (Passed or Flagged), the rules violated, the source, and a **Client** column that shows the client, for example ChatGPT.

### What to expect

* **Detection only.** ZeroDrift flags text and never rewrites it. When the assistant suggests wording, that wording is the assistant's, and ZeroDrift has not checked it until you validate it again.
* **Your workspace's active rules.** The result depends on the rules active in your ZeroDrift workspace.
* **A guardrail, not a security boundary.** ZeroDrift catches drafts that break your policies when the check runs. It does not stop a person, or an assistant with other tools, from sending content some other way.

### Example results

Wording from runs in ChatGPT:

* **Status:** "ZeroDrift: connected — signed in, 242 active rules". The number is your workspace's active rules. Status confirms you are signed in and validates nothing.

![ChatGPT showing ZeroDrift: connected — signed in, 242 active rules](https://mintcdn.com/zerodrift/WnYfzPOJ4zWh4vgY/images/openai-plugin/chatgpt-status.png?fit=max\&auto=format\&n=WnYfzPOJ4zWh4vgY\&q=85\&s=acb7982e002dab5dc9ff8171e65e1fac)

* **Permission prompt:** before it sends text to ZeroDrift, ChatGPT asks "Allow ChatGPT to use ZeroDrift?", lists the data it shares, and offers **Deny** or **Allow once**.
* **Passed:** "ZeroDrift: Passed — no issues found."
* **Flagged:** "ZeroDrift: Flagged — fix before sending (13 issues on 1 line: 12 do not send, 1 caution)". It is followed by the line and a table of severity, rule and policy. Severities shown are **Do not send** and **Caution**.
* **Several emails at once:** asked to validate the last five emails from Gmail, ChatGPT returned five Passed verdicts, each with a ZeroDrift job ID. ChatGPT noted that it checked subjects and message text, with long messages condensed, and did not extract attachment contents for the check.

### Verdicts

| Verdict | Meaning |
| - | - |
| **Passed** | The message goes through. |
| **Warned** | The message goes through. The assistant shows the warnings with the send. |
| **Blocked** | Do-not-send finding in a check before sending. The action is stopped. Hooks version only. |
| **Flagged** | Do-not-send finding in an on-request or skill-directed check. The assistant is told to withhold the send. Nothing technical stops it. |

Passed and Warned both permit the send, and Warned is shown with the send.

## Pick a distribution

Most people should use the marketplace install above. There is also an open-source hooks version of the plugin, which you only need if you want hooks. The two are not the same product, and the steps, enforcement and troubleshooting differ. Each section below names the distribution it applies to.

| | Hooks version (open-source repo) | OpenAI directory plugin (no hooks) |
| - | - | - |
| Source | [`ZeroDriftAI/zerodrift-codex-plugin`](https://github.com/ZeroDriftAI/zerodrift-codex-plugin) open-source repo, added as a Codex marketplace | The [ChatGPT plugin directory](https://chatgpt.com/plugins/plugin_asdk_app_6af4ea9184388191b9ee1ff948e85d34), shared by ChatGPT and Codex (package version 0.1.9) |
| Includes hooks | Yes | No. The package bundles and declares no hooks. |
| Checks before sending | Yes, enforced by hooks (Gmail) | Yes, directed by a skill. See [what each distribution enforces](#what-each-distribution-enforces). |
| Runs in | Codex CLI and Codex desktop app, per the repository README | Codex and ChatGPT, per the listing. See [what was tested](#what-was-tested). |
| Regions | Not specified | All regions. The listing is public to all. |

### What was tested

ZeroDrift tested the directory plugin in ChatGPT chat and ChatGPT Work, and confirmed the checks appear on the Activity page in ZeroDrift Command. If Codex asks you to sign in, `codex mcp login zerodrift` is the fallback.

OpenAI supports plugins in Chat and Work across ChatGPT on the web, desktop and mobile, and in Codex in the ChatGPT desktop app. The IDE extension does not support plugins ([supported surfaces](https://learn.chatgpt.com/docs/plugins)). That is OpenAI's platform support, not a statement of what ZeroDrift tested.

## What each distribution enforces

ZeroDrift is a guardrail, not a security boundary (see [How it works](#how-it-works)). There are three levels:

* **Enforced.** ZeroDrift stops the action when the verdict is **Blocked**. This needs hooks.
* **Skill-directed.** The plugin includes instructions that tell the assistant to check the message first and to withhold the send when the check flags it or cannot run. ZeroDrift does not technically stop the send. It depends on the assistant following the instructions.
* **On request.** You ask the assistant to check a draft and it returns the result. Nothing is stopped.

| Action | Hooks version (open-source repo) | OpenAI directory plugin (no hooks; Codex and ChatGPT) |
| - | - | - |
| Gmail send, draft, draft update, forward in Codex | **Enforced** before the action runs. A `Blocked` message is stopped. `Warned` and `Passed` go through. | **Skill-directed.** The skill tells the assistant to validate the final email, subject and body, before sending. |
| Slack messages | Not checked automatically. On request only. | **Skill-directed.** The skill tells the assistant to validate the final message before sending. |
| Any draft you ask the assistant to check (emails, messages, posts, documents) | **On request.** The result is returned; nothing is stopped. | **On request.** The result is returned; nothing is stopped. |
| Result is `Passed` or `Warned` | Goes through. | The skill tells the assistant to proceed and show any warnings with the send. |
| Result is `Flagged` | `Blocked` in a pre-send check: the action is stopped. | The skill tells the assistant to withhold the send and let you decide. |
| A message that cannot be checked (not signed in, service unreachable, validation unavailable) | Goes through. | The skill tells the assistant to withhold the send and report that it was not checked. |
| Hooks not trusted | Goes through. Nothing is checked before sending. | Not applicable. No hooks. |
| Technical send interception | Yes, for the Codex Gmail tools the hooks match. | **None.** The package supplies none. |
| Attachments | Not checked by the hooks. See [Limits](#limits). | Text, not files. See [Limits](#limits). |
| Original message of a forward | Not checked | Not checked |
| Saved drafts sent as-is | Not checked | Not checked |
| Sending or rewriting | ZeroDrift does neither. Sending needs another connected tool. | ZeroDrift does neither. Sending needs another connected tool. |

The directory plugin's pre-send check is **skill-directed validation. The assistant is instructed to withhold flagged or unchecked sends. Nothing at runtime enforces the block.**

The directory plugin's behavior was tested in ChatGPT chat and ChatGPT Work, and the checks behaved as described above: a flagged send is withheld, and a clean send goes through after the check passes. If the service is unreachable, the directory plugin withholds the send, and the hooks version lets it through. It is still a skill instruction, so treat it as a guardrail rather than a guarantee.

### Why the directory plugin has no hooks

Per OpenAI's documentation, lifecycle hooks run in the Codex runtime and need local setup and trust. A public directory package that contains lifecycle hooks is not currently eligible. See [OpenAI's plugin build guide](https://developers.openai.com/plugins/build/plugins#bundled-mcp-servers-and-lifecycle-hooks). That is why the directory plugin relies on a skill and the hooks version relies on hooks.

### ChatGPT Chat and ChatGPT Work

The directory plugin has no hooks in either. OpenAI's documentation says plugin hooks are not supported in cloud-orchestrated ChatGPT Work, and that synced Work supports only admin-defined MCP hooks ([current limitation](https://learn.chatgpt.com/docs/plugins)). Nothing establishes a hook mechanism for ordinary ChatGPT Chat either. In both, the directory plugin's pre-send check is skill-directed only.

An admin policy that allows only managed hooks makes Codex skip the hooks version's own hooks ([hook policy](https://learn.chatgpt.com/docs/hooks)). Managed hooks are a separate deployment option. They are not a prerequisite for the directory plugin.

## Hooks version

The hooks version is not a separate GitHub product. The [open-source repository](https://github.com/ZeroDriftAI/zerodrift-codex-plugin) is where the hooks version of the plugin lives, for Codex users who need hooks. Most people do not need it.

Applies to the hooks version only. It includes hooks.

**1. Add the marketplace and install**

```bash theme={null}
codex plugin marketplace add ZeroDriftAI/zerodrift-codex-plugin
codex plugin add zerodrift@zerodrift
```

**2. Sign in to ZeroDrift**

```bash theme={null}
codex mcp login zerodrift
```

In the browser, sign in to ZeroDrift and approve access for your workspace. No API key to copy.

**3. Trust the hooks**

In Codex, run `/hooks` and trust the ZeroDrift hooks. Codex asks again when an update changes them.

Until you trust the hooks, Gmail actions are not checked before they run.

## Limits

* **Only the text the assistant sends is checked.** The assistant decides what to send and formats it for the validation API, so it may condense long messages or leave out parts such as promotional links. ZeroDrift checks the text it receives. Anything the assistant leaves out is not checked.
* **Attachments.** ZeroDrift validates text, not files directly. To check an attachment, ask ChatGPT to extract and format its contents for validation. Only the text ChatGPT sends is checked.
* **Rate limit.** 50 requests per minute for a single validation.

## Data sent

| | |
| - | - |
| **What** | The text being checked (subject and body for email). Recipients are not sent. Attachments are not sent as files; see [Limits](#limits). |
| **Where** | `https://mcp.zerodrift.ai` over HTTPS, under your ZeroDrift sign-in. ZeroDrift records each check under your user and workspace. |
| **Retention** | See [Data retention](https://docs.zerodrift.com/data-retention.md). |

The body of a Gmail action must be plain text. If a body is multipart or base64, the hooks version's hook asks Codex to resend it as plain text before the check runs. The hooks also require a `subject` field, which can be an empty string.

## Troubleshoot

| What you see | What to do |
| - | - |
| The assistant says ZeroDrift is unavailable or asks you to sign in | Sign in again. In Codex, run `codex mcp login zerodrift` in a terminal. Never paste an API key or token into the chat. |
| Gmail actions go through with no check (hooks version) | Run `/hooks` in Codex and confirm the ZeroDrift hooks are trusted. An update that changes the hooks needs trusting again. |
| The assistant sent a message that was flagged (directory plugin) | The pre-send check is skill-directed, so the assistant can ignore it. Check Activity for the result, and ask the assistant to run the check again before sending. |
| A message went through and no check shows in Activity | The check did not run. Sign in again, confirm the plugin is installed from the directory (or the hooks are trusted for the hooks version), then retry. |
| Codex asks you to resend the message as plain text, or to add a subject | The hook can only check a plain-text body and a subject field. Follow the request. |
| You cannot approve access in the browser | Sign-in uses OAuth. OpenAI has no separate approval workflow. A workspace admin creates and authorizes a shared connection, and each user approves the permissions the plugin requests as a separate step ([shared connections](https://learn.chatgpt.com/docs/enterprise/shared-connections)). See [Before you start](#before-you-start) for who can approve access in ZeroDrift. |

## Disconnect

* **Hooks version:** run `codex mcp logout zerodrift` to sign out. Uninstall the bundle separately through `/plugins`.
* **Codex, either distribution:** `codex mcp logout zerodrift` removes the stored sign-in. `codex mcp remove zerodrift` deletes the stored server definition. `codex plugin remove` removes the plugin from local configuration and cache. Command names come from [OpenAI's developer commands reference](https://learn.chatgpt.com/docs/developer-commands?surface=cli).
* **Directory plugin in ChatGPT:** open **Plugins > ZeroDrift**, select the **...** menu next to **Settings** and **Try in chat**, and choose **Uninstall**. The menu also has **Copy link**. Uninstalling removes only the bundle. A separately connected MCP integration stays connected until you disconnect it in ChatGPT. A workspace-installed or default plugin may not show **Uninstall**, because an admin controls it. See [removal guidance](https://learn.chatgpt.com/docs/plugins#remove-a-plugin).
* **ZeroDrift Command:** open **Connections**. Under **Enforcement surfaces**, the **ChatGPT and Codex** card shows a **Connected** status, when it was last used, and a **Manage** button. **Manage** opens a **ChatGPT and Codex** dialog. Its **Use in Codex** section shows `codex mcp login zerodrift` ("In a terminal: sign in to ZeroDrift") and `/hooks` ("In Codex: trust the ZeroDrift hooks"). Its **Connection** section shows "Connected since" with the date and a red **Disconnect** button. The `/hooks` line applies only to the GitHub repo plugin. The directory plugin has no hooks. **Disconnect** ends the connection. Select **Done** to close the dialog.
* **Workspace admins:** set the plugin's installation policy to **Disable**. Disabling does not uninstall the plugin or disconnect existing connections, so review role overrides and revoke the provider authorization separately ([admin controls](https://learn.chatgpt.com/docs/enterprise/apps-and-connectors)).

## Next steps

> [card] [Check an Email](https://docs.zerodrift.com/guides/openai-plugin/validate-email.md)
>
> Ask the assistant to check an email. See Flagged, Passed, and suggested wording.

> [card] [Check a Social Post](https://docs.zerodrift.com/guides/openai-plugin/validate-social-post.md)
>
> Check a post for X before you publish it.

> [card] [Pre-Send Email Check](https://docs.zerodrift.com/guides/openai-plugin/pre-send-email.md)
>
> Have the assistant check an email before it sends. The hooks version and the directory plugin, in separate sections.

> [card] [Batch Check and Activity](https://docs.zerodrift.com/guides/openai-plugin/batch-check-and-activity.md)
>
> Check your last five emails, then review each check in Activity.

## Support

[support@zerodrift.ai](mailto:support@zerodrift.ai)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.