AI Agents

What Is a Plugin, and Why Won't It Let My AI Agent Do Everything?

The word shows up in Codex and Claude Code, nobody explains it, and your agent still can't do half of what you asked. Here is what a plugin really is, and why one is never enough.

Short answer: A plugin is a single connection that gives your agent access to one outside tool, like your email or your store. It is not a power switch that unlocks everything. Your agent can't do it all because each capability is a separate door, and most "it won't do that" moments are really "it doesn't have the key yet."

You asked your AI agent to do something completely reasonable. Follow up with everyone who registered for last week's webinar. Send the invoice. Pull this month's numbers from your store. And instead of doing it, the agent gave you some version of "Gmail is unavailable" or "I don't have access to that."

You expected a genie. You got a polite shrug. And somewhere in that exchange a word appeared, plugin, or connector, or integration, and now you are not even sure what you were supposed to turn on.

Here is the plain-English version, and why the answer is simpler than the jargon makes it sound.

What a plugin actually is

A plugin is a connection. That is the entire concept. It plugs your agent into one outside thing: one app, one account, or one specific skill. Gmail is one. Your calendar is another. Google Drive, your CRM, your online store, your accounting tool, each of those is its own separate connection.

Depending on which tool you use, you will see this called by a different name. In Codex it is a "plugin," listed under a "Plugins" section. ChatGPT usually says "connector," and Claude Code says "plugin" or "MCP server." Other tools say "integration" or "app." The label keeps changing. The idea never does: one connection to one outside thing.

Before you connect anything, picture your agent as a brilliant new hire on their first morning, standing in an empty office. Sharp, fast, ready to work, and completely locked out. No email login, no keys, no files. It can't read your inbox, because nobody has handed it your inbox. A plugin is you handing over one key.

Why one plugin doesn't unlock everything

So you connect Gmail. Now the agent can read and send email. Wonderful. Then you ask it to post that same message to Instagram, and you are right back to the shrug. This is the part that frustrates people, so let me be specific. There are three reasons.

First, each plugin is one door. The Gmail connection does email, and only email. It does nothing for Instagram, your calendar, or your store. Those are three more doors, and each one needs its own key. Ten jobs across ten tools means ten connections. There is no single plugin that does everything, the same way there is no single key that opens every building in your city.

Second, access is set separately from the plugin, and it usually starts locked down. You can connect Gmail and still only allow it to read, not send. Then you ask it to send the follow-ups, it can't, and it looks broken. It isn't. You gave it a key to the room and told it not to touch anything. A surprising share of "my agent won't do that" is really "I connected it, but left it on read-only."

Third, a plugin only offers the specific actions its builder wired up. Open one in Codex and you will see it bundles a named list of "skills," plus a "capabilities" line spelling out what it can actually do. A connection to your accounting tool might let the agent read transactions but not delete them, on purpose. The agent can't invent an action the plugin doesn't expose. It isn't being lazy. That capability was simply never built in.

So when your agent stops short, it is almost always one of these:

When the agent won't do itThe real reasonWhat actually fixes it
"Gmail is unavailable"The tool isn't connected yetAdd the connector for that tool
It reads your email but won't sendConnected, but read-onlyGrant send or write access
It won't touch your CRM at allNo plugin exists for that toolBrowser control, or a custom connection
It refuses to delete or overwriteA safety guardrail, or that action was never builtConfirm it explicitly, or accept the boundary

If you have hit the first row of that table inside Codex, the walk-through in how to fix "Gmail is unavailable" in Codex is the exact fix.

Stop picturing a genie. Picture a new hire.

This is the reframe that makes the whole thing click. A genie grants any wish out of thin air. That is not what you have. You have a new hire.

On someone's first day, you do not hand them "everything." You would never. You give them the email login. Then a seat in the CRM. Then the shared drive, then the calendar, each at the right level of access, each when the work actually calls for it. Six months later they can run half your operation, not because they suddenly got more talented, but because they accumulated access one credential at a time.

Your agent works exactly the same way. The plugins are the credentials. So the job was never to find the one magic plugin that does everything, because it does not exist. The job is to decide which rooms a task actually needs, and hand over those specific keys. Some tools are far easier to hand keys to than others, which is worth knowing before you pick where to start.

How to actually get your agent to do the thing

So the next time your agent stops short, walk through this instead of assuming it is broken.

  1. Name the tool the task actually touches. "Follow up with registrants" touches your email, and probably your registration list or CRM. Get specific about which apps are involved before you do anything else.
  2. Connect that one thing. Add the connector or plugin for it. In Codex that lives under Plugins in the workspace dropdown, and here is the exact click-by-click. If you are not sure how, that is a fair thing to make the agent itself walk you through, one step at a time. When it tells you a tool is unavailable, that is usually the exact thing to fix.
  3. Grant the right level of access. If you want it to act and not just look, you have to allow it to send, write, or post, not only read. Read-only is the single most common reason a connected agent still does nothing.
  4. If no plugin exists for your tool, you still have two paths. You can let the agent drive your web browser directly, clicking and typing on the screen the way you would, which works even when there is no formal connection. Or someone can set up a custom connection for that specific tool.

One quiet rule sits underneath all of it: an agent can only use a door it knows it has. After you connect something, say so. "You now have access to my Gmail and my registration sheet. Use them to send the follow-ups." A connection plus a clear instruction is what turns the shrug into finished work.

Quick answers

What is a plugin for an AI agent? A single connection that gives your agent access to one outside app or one new skill, like your email, calendar, or store. It is not a switch that unlocks every capability at once.

What is the difference between a plugin and a connector? Mostly just the name. ChatGPT usually says connector, Codex says plugin, and Claude Code says plugin or MCP server. All of them mean one connection to one outside tool.

Why won't my agent do everything I ask? Usually the tool isn't connected, it is connected but read-only, or the action you want was never built into that connector. Each capability is a separate connection with its own permission level.

How do I let my agent use a tool that has no plugin? Let it control your browser directly, or have someone build a custom connection for that tool. Then tell the agent it now has access.

Your agent was never going to do everything on its first day. Neither would the sharpest person you have ever hired.

Stop hunting for the plugin that does everything. Decide which rooms the job needs, hand over those keys, and tell your agent to get to work.


Back to all articles