Plugins and connected services

Back to the Codex learning hub

Goal and preparation

Lessons and resources mentioned here: Permissions and data scope

Plugins bundle Skills, MCP and supporting components. Skills describe workflows; MCP tools access services. Installation, service authentication and tool availability are separate states: an installed plugin may still be unable to read a document.

Step 1: Confirm your surface

Open Plugins and inspect the OpenAI, workspace, Personal and Installed categories shown. On Windows/macOS/Linux, check app version and plugin requirements; confirm Linux-preview availability locally. A public listing does not prove support. See Platform selection for desktop setup.

On Windows/macOS/Linux, enter your practice folder in a terminal, run codex, then enter /plugins inside Codex. The first is a shell command, the second an interactive command. Plugins are unsupported in the IDE extension; use desktop or CLI.

codex

/plugins

A mobile plugin does not automatically expose a computer's disk. Tasks needing local computer files still follow Mobile and cross-device connection requirements; this exercise reads only the chosen service document. A Desktop only plugin cannot run directly on a phone merely because its listing is discoverable on the web.

Step 2: Prepare a verifiable document

In your practice Google Drive, create a document titled Codex Plugin Practice and paste the full text below. Keep it private; do not enable public access merely to make a plugin read it. Retain its URL to specify scope. If using another service, first complete installation and state recording, then adapt the example only after confirming equivalent read tools.

Codex Plugin Practice
Revision: 1
Document marker: PLUGIN-LAB-A
Task count: 3
Completed count: 2
Pending count: 1
Scope: fictional practice data; read only.

The marker and revision identify what the tool actually read, rather than testing whether a model can guess three counts. Titles can repeat; the URL and marker must match. If search finds two documents with the same name, do not choose arbitrarily. Inspect the original service page first, then request that exact URL for an independent comparison.

Step 3: Install, connect and start a fresh task

Open plugin details and verify the name, provider, marketplace source and capabilities before installing the intended Google Drive plugin. Desktop documentation describes a plus button; CLI installation is available from details under /plugins marketplace categories. Record the name or ID and date so a similarly named package is not mistaken for the same installation.

When sign-in is required, verify the service and account own the practice document, then read the requested permissions. Do not paste passwords, verification codes, tokens or connection callback URLs into chat. If the requested access exceeds your acceptable purpose, cancel and record not connected. A read-only prompt does not narrow OAuth permissions.

Return to the original surface, inspect Installed and connection state, then start a new chat or CLI session. Some plugins connect only on first use, so installation alone cannot skip this check. API-key Codex authentication does not support every OAuth-dependent plugin. Check authentication mode and plugin requirements rather than repeatedly replacing or printing keys.

Step 4: Verify an actual read

On desktop, use @ to select the installed plugin or bundled skill and paste the request, replacing the marked position with your practice document URL. In a new CLI session, request the installed plugin for the same document and use the skill-selection entry if shown. Read only the specified document, making both results and scope easy to verify.

Use the installed Google Drive plugin to read this practice document:
<PASTE YOUR PRACTICE DOCUMENT URL HERE>
Return the document title, source link, revision, marker, and the three counts.
Use an actual read result. If access is unavailable, say which step failed.
Do not edit, share, move, delete, or search unrelated files.

The reply should link the same document and report Revision 1, PLUGIN-LAB-A and 3/2/1 with actual read evidence. Match values to total, completed and pending fields rather than merely checking the numbers. Then manually change Revision to 2, Completed to 1 and Pending to 2. Request a fresh read without supplying the new values and verify it reports the updated document.

If it returns old counts, request a fresh source read and observed revision. Record connection, account-access or update-delay errors instead of fabricating new results. A fresh task can help separate old conversation context. Finally inspect the service to confirm the agent did not edit, share or move the document, distinguishing your manual revision from the agent's read.

Add a marker the model has not been told

After the two fixed-count checks, change Revision to 3 in the source service and replace Document marker with a new non-sensitive marker of your own; keep the second round's counts at 3/1/2. Put the marker only in the document, not in the Codex prompt. Start a new task with the same document URL and earlier read request. Count this round as verified only when a real tool result supplies your new marker. Repeating the article's marker or the previous revision does not pass.

Fill the record below in your own notes; keep private URLs private and use aliases in public reports. Installed records the bundle, Connected the service connection, and Fresh read needs its own read time, source and marker evidence. After uninstalling, a missing skill does not automatically make Connected false. Check separately connected MCP integrations using the official plugin documentation, then handle only the authorization created for this exercise if needed.

# Plugin read verification
Surface / host / date:
Plugin ID and source:
Practice document alias:
Installed: <observed state>
Connected: <observed state or unconfirmed>
Tool available in this new task: <evidence>
Fresh read time and source: <tool evidence or NOT RUN>
Revision and private practice marker: <actual result, do not copy the tutorial value>
Source unchanged by the tool: <verification>
Bundle removed: <observed or NOT RUN>
Separate connection disconnected: <observed or NOT RUN>

Step 5: Disable, uninstall and disconnect

In CLI /plugins, Space toggles an installed entry; Uninstall plugin in details removes it. On desktop/web, open installed plugin details for that action. Workspace/default plugins may be administrator-controlled. Record restrictions; do not delete app caches or installation folders.

Uninstalling a bundle and disconnecting an external account are separate. Independently connected MCP integrations may remain; disconnect the service in ChatGPT connection management and inspect the provider's third-party authorization page if needed. Check shared use before removing only the practice connection. In a fresh task verify that connection can no longer read, rather than treating a removed Installed entry as proof all access was revoked.

Record installation, authorization, first read, revised read and removal separately. Mark unoperated platforms/services as documentation-checked. This lesson does not claim to have authenticated your Drive or read private documents. Diagram 1 selects capability/scope, 2 connects and reads, and 3 verifies and closes. Continue to Plugin troubleshooting for layered diagnosis.

Back to the Codex learning hub

Read the full description

Plugin to Connect to Tool

• AI tools
• Tutorials
• Codex