ChatGPT custom connectors (MCP): how to add one, and the same in Claude

What a custom connector is

A custom connector links your AI to an outside server that gives it new tools. ChatGPT and Claude come with ready-made apps for big names such as Google Drive or Slack. A custom connector is how you add anything else: a tool your own company runs, or a server someone else built for a service you use.

Both AIs talk to these servers over MCP, the Model Context Protocol. It’s an open standard for how an AI asks a server which tools it has and then uses them. You don’t need to understand it. What you need from whoever runs the server is its address, a web link that usually ends in /mcp, and a way to sign in.

The names keep changing. OpenAI’s older help pages say connectors, later ones say apps, and in today’s ChatGPT the button reads “Add custom MCP server” and the result shows up as a plugin. Claude still says custom connector. People search for “ChatGPT custom connector”, so that’s the name we use here, for both.

Which ChatGPT plans can add a custom connector

Guides online disagree on this, partly because OpenAI has changed the rules more than once. We went by OpenAI’s own article, Connecting custom MCP servers in ChatGPT, as it read on 31 October 2026, a few days after its last update. Claude’s rules come further down, for comparison.

Three more points from the same article:

• Custom MCP servers are added on the web; OpenAI says it can’t be done from the mobile app.
• You don’t need Developer Mode any more. Older guides tell you to switch it on first; now ChatGPT asks you to accept a risk warning for each server.
• Deep research can use a custom connector to read, never to change anything.

If your plan isn’t in the table, Claude is the easier way in. Its free plan already allows one custom connector.

How to connect a custom connector in ChatGPT

These steps follow OpenAI’s article, and they match our setup guide, where APIUniCon is the example. Claude’s version is in the next section. On a computer, at chatgpt.com:

1. Open Plugins at chatgpt.com/plugins.
2. Click “+” and choose “Add custom MCP server”.
3. Give it a name and, if you like, a description. Under Connection, paste the server’s address. For APIUniCon that’s https://apiunicon.online/mcp.
4. Set up the sign-in. For APIUniCon choose OAuth: you sign in on our page, so the AI never gets your password.
5. Read the risk warning, tick “I understand and want to continue” and click “Create as a plugin”. You’ll see the tools the server offers.
6. Install the plugin and sign in. Then start a new chat, type @ and pick the plugin from the list.

In a Business workspace an admin or owner goes through these steps, tests the plugin and publishes it, and the team then finds it under Settings → Plugins. One detail from OpenAI is easy to miss. After an admin approves a server, the workspace keeps a snapshot of its tools. If the server adds tools later, they appear only after an admin refreshes the list.

How to add a custom connector in Claude

Anthropic’s help article, Get started with custom connectors using remote MCP, lists custom connectors on the Free, Pro, Max, Ultra, Team and Enterprise plans. The free plan gets one. On Team, only owners and primary owners add a connector, and each member then connects to it. On claude.ai or in the desktop app:

1. Open Customize → Connectors (claude.ai/customize/connectors).
2. Click “+ Add” and choose “Add custom connector”.
3. Enter a name and the server’s address, then click “Continue”.
4. Claude shows the sign-in settings it found. Click “Continue” again, choose “Sign in now”, keep “Use Claude’s published identity” for the client, as Claude suggests, then click “Add”.
5. Sign in on the server’s own page.
6. In a chat, click “+” → Connectors and switch the connector on.

Side by side:

Keys, permissions and changes: the safety side

A custom connector can read your data and, on some plans, change it. OpenAI and Anthropic give the same advice in their help articles: connect only servers you trust. A bad server, or a web page with hidden instructions, can try to steer the AI; this is called prompt injection. Four habits help, in ChatGPT and Claude alike:

• Keep keys and passwords out of the chat. A server that needs your service’s key should take it on its own page. With APIUniCon the AI sends you a link, you paste the key there, and we keep it encrypted. Neither ChatGPT nor Claude ever sees it.
• Sign in on the server’s own page. Choose OAuth where the server offers it: you type your password on the server’s site, and the AI never handles it.
• Leave the confirmation questions on. ChatGPT lets you pick when it asks before using an app: “Always ask”, “Allow read actions”, “Allow low-risk actions” or, for some apps, “Allow all actions”, which OpenAI marks as riskier. Claude asks before it uses a connector’s tool; click “Allow always” only for tools you’d let run unsupervised.
• Give each service key only the rights it needs. For answers alone, a read-only key is enough.

APIUniCon adds its own layer on top. The AI can read on every plan, and changes are on Pro and Business. Before each change it shows exactly what it’s about to do and waits for your “yes”, and deleting anything always goes through our page. Every change is recorded in the change log of your APIUniCon account.

One custom connector for every service with an API

Most custom connectors serve one service. APIUniCon works the other way round: you add it once and then connect your services inside it, from the chat. Give the AI a link to a service’s developer docs, create a key in that service’s settings and paste it on our page. From then on you just ask.

This matters most on Claude’s free plan, which has room for one custom connector. With APIUniCon in that slot, every service you connect fits through it. Our guide to connecting any service with an API walks through the whole path on one example.

What to ask once it’s connected

Start with questions that show what the connector can reach, then move to your data:

• Which services are connected through APIUniCon right now?
• Connect my Kommo account.
• Which of the tools you got from APIUniCon only read, and which can change data?
• Which Allegro offers will run out within 7 days at the pace of the last 30 days?
• Which Kommo deals have been sitting in Negotiation for more than 14 days?
• What follow-ups are overdue in Less Annoying CRM, and for which contacts?
• How do this month’s Allegro sales compare with last month’s?
• Before you change anything, show me exactly what you’re going to send.

Cases: three setups and what happens in each

The people below are made up. The plan rules they run into are OpenAI’s, as of 31 October 2026.

“I’m on Plus and there’s no Add custom MCP server”

A shop owner pays for ChatGPT Plus, opens Plugins and can’t find the button. OpenAI’s article leaves Plus out of the list, so this isn’t a mistake on their screen. They have two ways forward. They can add APIUniCon to Claude, where the free plan allows one custom connector, and keep ChatGPT for everything else. Or, if the business already has a ChatGPT Business workspace, the admin can add it there.

A Business workspace where only the admin can add it

An operations manager wants the whole team to be able to ask about orders. In a Business workspace only admins and owners create custom MCP servers, so the manager asks the admin. The admin adds APIUniCon with the six steps above, tests a couple of questions and publishes it. The team finds it under Settings → Plugins. If APIUniCon gets new tools later, they show up after the admin refreshes the list in workspace settings, as the snapshot rule above describes.

Pro: answers yes, changes no

A consultant works in ChatGPT Pro rather than Claude and adds APIUniCon to keep an eye on a CRM. Questions are what Pro allows: overdue tasks, stalled deals, last month’s totals. Moving a deal to the next stage is a change, and on Pro OpenAI gives custom MCP servers read and fetch permissions only. For changes the consultant needs a Business, Enterprise or Edu workspace, or Claude, plus a Pro or Business plan on our side.

What the AI can do through a custom connector

The tools come from the server, so a custom connector can do as much as the server allows and your plan permits. Through APIUniCon the AI can also change data in the services you connect, and each change waits for your confirmation:

• “Set the mug set’s stock on Allegro to 40.” First you see the offer with its old and new quantity.
• “Reopen the Brightline deal in Kommo and move it to Negotiation.” Nothing moves until you’ve seen the deal, its current stage and the new one.
• “Add a follow-up call with Sarah Jones for Thursday.” The task, the contact and the date come up, and the task is added once you agree.
• “Delete the test contact I made yesterday.” This one opens our page. Deleting always does.

If something goes wrong

• There’s no “Add custom MCP server” button. Check your plan against the table above. In a workspace, an admin may also have turned off the “Create plugins with MCPs” permission.
• APIUniCon doesn’t show up in the chat. In Claude, click “+” → Connectors and check that it’s on. In ChatGPT, type @ and pick it from the list.
• Claude’s free plan already has a custom connector. Open Customize → Connectors, remove the one you don’t need and add APIUniCon. The setup guide covers this and a few other snags.