ChatGPT MCP lets ChatGPT use tools and data from other systems through the Model Context Protocol. There are now three ways to connect a server: as a plugin on the web, in the new desktop app, or through the OpenAI API. This guide explains each route, which plans can use it, the setup steps, and what usually breaks.
ChatGPT MCP: the short answer
- On the web, you add a remote MCP server as a plugin. Business, Enterprise and Edu get full support, while Pro gets read-only access.
- In the ChatGPT desktop app, you add local or remote servers under Settings, and Codex uses the same setup.
- Through the API, developers pass an MCP server to the Responses API as a tool.
Does ChatGPT support MCP?
Yes. MCP is an open standard that lets AI apps call tools on other systems in the same way, whatever the app. If you are new to it, our introduction to Claude MCP explains the idea with examples.
For the technical details, the official specification is the place to start.
ChatGPT first added full MCP support through a beta called Developer Mode in September 2025. Since then, OpenAI has changed the setup twice. First, it moved custom servers into Plugins, so Developer Mode is no longer needed. After that, it launched a new desktop app that can run MCP servers on your own computer. As a result, many guides and videos still show steps that no longer exist.
Which ChatGPT MCP route fits you
The right route depends on who you are and where the server runs. In short, this table sums it up.
By contrast, Free, Go and Plus users cannot add a custom MCP server on the web. However, they can still use the apps in ChatGPT’s plugin directory. Our guide to ChatGPT connectors lists the most useful ones and what each plan gets.
How to add a custom MCP server in ChatGPT on the web
Who can do it
On Business, only workspace admins and owners can add a custom MCP server and publish it for the team. On Enterprise and Edu, admins can also give selected members permission through roles. Pro users can add a server for themselves, but ChatGPT only allows read and fetch actions from it. In every case, this works on the web, not in the mobile apps.
The steps
- In ChatGPT on the web, open Plugins, select the + button, and choose Add custom MCP server.
- Next, enter a name and, if you like, a description.
- Then paste the server’s public HTTPS address, or pick a Secure MCP Tunnel for a private server.
- Choose how it signs in. Most servers use OAuth.
- Tick the box that confirms you understand the risks, then select Create as a plugin.
- Finally, review the tools ChatGPT found, then sign in to the server if it asks.
After that, start a new chat, type @, and pick the plugin to test it. On Business and Enterprise, an admin then publishes it, so others in the workspace can use it. OpenAI’s guide to connecting custom MCP servers has the full details, including the older Apps settings that some workspaces still use.
After it is published
ChatGPT keeps a fixed snapshot of the server’s tools once an admin approves it. Therefore, when the server’s developer adds or changes a tool, nothing changes in ChatGPT until an admin refreshes the tool list. On Enterprise and Edu, admins can also switch individual actions on or off.
How to use MCP in the ChatGPT desktop app
The new ChatGPT desktop app for macOS and Windows brings Chat, Work and Codex together in one app. It can connect to MCP servers directly, including servers that run on your own computer. Moreover, it shares its MCP setup with the Codex command line tool and the Codex IDE extension. So you only set up a server once.
Add a server in the app
- Open Settings, then select MCP servers.
- Select Add server.
- Enter a name, choose STDIO for a local server or Streamable HTTP for a remote one, and give the command or the address.
- After that, save it and select Restart.
- If the server uses OAuth, select Authenticate and sign in.
In a chat, type /mcp to see which servers are connected. OpenAI’s MCP guide for ChatGPT and Codex lists every option. OpenAI’s page does not say which plans include this, so check your plan before you rely on it.
Or edit the config file
Alternatively, the same settings live in a text file, by default ~/.codex/config.toml. Each server gets its own block. For example, this adds one remote server and one local server:
[mcp_servers.docs] url = "https://mcp.example.com/mcp" [mcp_servers.files] command = "npx" args = ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/projects"]
You can also decide when ChatGPT asks before using a tool. For instance, the “writes” approval mode asks only for tools that are not marked as read-only.
How to use MCP with the OpenAI API
Developers can also skip ChatGPT entirely and call an MCP server from their own code. In the Responses API, an MCP server is simply another tool. Here is OpenAI’s own example, which uses its public documentation server:
from openai import OpenAI
client = OpenAI()
resp = client.responses.create(
model="gpt-6-astra",
tools=[
{
"type": "mcp",
"server_label": "openai_docs",
"server_description": "Search and read the public OpenAI documentation.",
"server_url": "https://developers.openai.com/mcp",
"require_approval": "never",
}
],
input="Search the OpenAI docs for Responses API streaming and return the relevant links.",
)
print(resp.output_text)
For servers that need a login, you pass an OAuth access token in the authorization field. In addition, the allowed_tools field limits which tools the model can see, which saves cost on large servers. Keep approvals on whenever the tools can change data. The details are in OpenAI’s MCP servers guide for the API.
Build your own: Plugin Creator and Sites
If no server exists for a system you use, ChatGPT can now help you build one. You start a Work task with Plugin Creator, attach the system’s API documentation, and describe what you need. Then ChatGPT Sites can host the MCP server, so you do not need your own hosting.
In general, this suits simple, internal tools, such as looking up a ticket by number. For anything that many people rely on, a proper server is still the safer choice. Our guide to building an MCP server covers that route.
What breaks in practice
Most ChatGPT MCP problems look the same: the server connects but does not behave. These are the usual causes and fixes.
Sign-in problems are the most common of these. Our guide to MCP OAuth explains how the sign-in works and why tokens expire.
After that, rate limits are the most common, because ChatGPT can call a tool many times for one question. Our MCP rate limits guide explains how to avoid them.
ChatGPT or Claude for MCP?
Both apps support remote MCP servers with OAuth, so the same server usually works in both. However, there are differences. The main difference is who can add one. In Claude, custom connectors work on every plan, and the Free plan gets one. In ChatGPT, full custom servers on the web need a Business, Enterprise or Edu workspace.
Distribution also differs, for instance. ChatGPT has a large public plugin directory and admin controls for publishing apps to a team. Claude, by contrast, makes it simpler for one person to add a server for themselves. If your team uses both, pick servers that support standard OAuth, so each person can connect their own account in either app.
Where your data goes
When ChatGPT uses a custom MCP server on the web, OpenAI’s servers call it. The results then become part of the conversation. So two sets of terms apply: OpenAI’s, and those of whoever runs the MCP server. On Business, Enterprise and Edu, OpenAI does not use app data to train its models by default.
For European companies, therefore, the questions to ask are where the MCP server stores data and which law applies. Our overview of GDPR-compliant MCP servers covers the checks to make.
In addition, under the EU AI Act, some uses of AI on business data bring duties for the company using it. Our guide to the EU AI Act obligations for MCP explains which ones apply.
For the technical side, such as limiting what a server can reach, see our guide to MCP security.
Example: connecting Peliqan to ChatGPT
Peliqan runs one MCP server that covers many business apps at once, so ChatGPT can answer questions across them. For example, you can connect your CRM, accounting tool and webshop in Peliqan, then ask ChatGPT one question that combines all three. Our guide to cross-source SQL explains why that is hard with one server per app.
How to set it up
First, connect the apps you want to use in Peliqan. Then, in ChatGPT on the web, add a custom MCP server with the address https://mcp.peliqan.io/mcp and leave the sign-in on OAuth. Finally, log in with your Peliqan account and approve access. Peliqan registers ChatGPT’s sign-in address in advance, so you do not need to create a client ID. The Peliqan MCP server documentation covers the details.
What it can do
Questions run as SQL on a synced copy of your data in Peliqan’s data warehouse. As a result, one question can join several apps, and it does not hit each app’s API every time. Peliqan can also query some sources in place when you need live data.
When ChatGPT needs to change something, such as updating a contact, it uses writeback via MCP. Writeback is on by default, and you can switch it off per connection. Keep in mind that on a Pro plan, ChatGPT only allows read actions from custom servers, so writeback will not work there.
For example, CIC Hospitality brought data from more than 50 sources together in Peliqan and saves around 40 hours a month on reporting. You can read the full case study for more.
Build your own server instead
In addition, if you need custom tools, Peliqan can host your own MCP server, written as Python functions. The guide to building a remote MCP server on Peliqan walks through it.
For a finance example, see our build guide for EU accounting.
If you are comparing options first, however, our list of 13 MCP servers every European business should know is a good start.
Pricing is based on the number of data sources and the plan you choose; see the pricing page for details.
For a wider view of what MCP servers cost, read our guide to MCP server pricing.
Conclusion
ChatGPT MCP is no longer one beta feature behind Developer Mode. On the web, workspace admins add custom servers as plugins, and Pro users get read-only access. In the new desktop app, you can run local and remote servers next to Codex. And through the API, developers can use any remote server as a tool.
So start with the route that matches your plan. Then test with read-only questions, and turn on write actions only when you trust the results. To see how one server can cover all your business apps, visit the Model Context Protocol hub.
Alternatively, if you prefer a walkthrough with your own apps, book a demo.



