ChatGPT MCP support means you can paste the address of a remote MCP server into ChatGPT as a custom connector, and ChatGPT will then call that server's tools inside a chat. It works in the ChatGPT web app, and you need no code. This guide covers plan availability, where to click, the URL-key option, and what to check when it fails.
Which ChatGPT plans can add a custom MCP connector?
OpenAI's help article lists Business, Enterprise and Edu workspaces on the web; some guides also report Plus and Pro. Check Settings in your own account. OpenAI changes this list and the menu names from time to time, so your own account is the final word.
Workspace plans add a second gate. On Business, Enterprise and Education, an admin can restrict or switch off custom connectors, so the option may be missing for you even when your plan supports it. OpenAI's own page says that account and workspace policies apply to adding and using custom MCP servers.
- Personal accounts: check Settings for the connector or developer option. If it is not there, your plan or region may not have it.
- Workspace plans: ask your admin to allow custom connectors if the option is greyed out or absent.
Custom connectors are a web feature. Set the connector up in ChatGPT on the web. Whether the phone apps use it afterwards depends on your plan and OpenAI's current rollout. The developer-mode toggle itself, and its own rules, are covered in ChatGPT developer mode and MCP.
What do you need before you start?
You need three things: a ChatGPT account that can add custom connectors (see the plan note above), a remote MCP server with a public HTTPS address, and a way to sign in to it. OpenAI's guide asks for a public HTTPS endpoint, usually ending in /mcp, that speaks streamable HTTP. The remote MCP server entry explains the term, and what MCP is covers the basics if the acronym is new.
Running your own server means hosting, updating and securing it. A hosted gateway skips that. With MCPifex you pick a server in the marketplace, save your credentials, choose which tools are allowed, and get one address plus an API key. The free plan includes 3 instances and needs no card, see pricing.
How do you add an MCP server to ChatGPT, step by step?
These steps use an MCPifex instance, but the ChatGPT side is identical for any remote MCP server. OpenAI has renamed menus several times, so the names on your screen may differ slightly. The steps follow the path in OpenAI's current guide, which starts from ChatGPT Plugins.
- Prepare the server. In the MCPifex portal, open your instance, switch on only the tools you want, and run Test connection. Create an API key. Keep write tools off unless you need them.
- Copy the connection address. The address is the gateway URL with your key at the end:
https://mcp.mcpifex.com/mcp/<YOUR_MCPX_KEY> - Open the connector screen. In ChatGPT on the web, open Plugins, select the plus button, and choose Add custom MCP server. You may first need to turn on Developer mode in Settings under Security and login. Older accounts may instead show Settings, Connectors, Advanced, Add connector.
- Name it. Use a name you will recognise, such as "My PostgreSQL". Add a short description of what the server does.
- Paste the URL. Paste the full address, with the key in it, into the Connection field. It must be the public URL including
/mcp. - Choose No Authentication. ChatGPT may offer OAuth, No Authentication and sometimes Mixed. Pick No Authentication and leave the OAuth fields empty. The gateway still checks the key inside the URL.
- Accept the warning and create it. ChatGPT shows a risk notice about custom connectors. Read it, accept it, and select Create.
- Check the tool list. ChatGPT reads the server and lists the tools it found. Compare that list with the tools you enabled in the portal. Both should match.
The same flow is on the Connect ChatGPT page, and the ChatGPT docs keep the click-by-click version.
Why does the key go in the URL?
Web connectors in ChatGPT cannot send a custom header such as Authorization. Desktop and command-line clients can send one. So for ChatGPT, MCPifex accepts the key as the last part of the path, in the form /mcp/mcpx_.... The gateway reads the key from the path and checks it exactly as it would a header. Clients that can send headers use the plain gateway address and a bearer header instead, see API keys and authentication.
Treat this URL like a password. Anyone who has it can use the tools you enabled. Keep it out of screenshots, shared documents and chat prompts. Only the MCPifex key belongs in it. Database passwords and Google logins stay in the instance settings.
If the URL leaks, open the portal, revoke the key and create a new one. Then update the connector in ChatGPT with the new address. The MCP security guide covers more habits worth having.
How do you use the connector in a chat?
Start a new conversation and select the connector. In OpenAI's current flow you type @ in the prompt box and pick it. In older flows you open the plus menu, choose Developer mode and tick the connector. A connector you added mid-conversation may not show up until you start a new one.
Then ask for something small and checkable. With a PostgreSQL instance, try "List the tables in my database." Expand the tool call that ChatGPT shows and read the data it returned. A general explanation of what a table is does not prove the server ran. The portal's call log for the instance shows whether the request arrived.
ChatGPT asks for your confirmation before it sends data to a connector, and it can ask again for write actions. OpenAI's API documentation says approval is required by default for remote MCP servers. You can often choose to remember a decision inside one conversation, so a prompt may not appear every time. Read the input before you approve anything that changes data.
What works and what does not?
WorksDoes not, or varies
Calling tools from a remote MCP server over HTTPS
Servers that run only on your computer (stdio) cannot be added directly
Read tools such as listing tables or fetching report data
Write tools need your confirmation
Several connectors in one account
Large tool lists can make ChatGPT slower or choose the wrong tool
Changing enabled tools in the MCPifex portal
ChatGPT may need a refresh or a new chat to see changes
Web app
Availability in mobile and desktop apps follows OpenAI's rollout
MCP also asks a lot of the model. A server with 88 tools gives ChatGPT a long menu. On MCPifex you can allow only the tools a task needs, which keeps the list short and keeps risky tools out of reach. Write and destructive tools start switched off, and tools you turn off are not offered to the AI.
Developers using the API have a different route. The Responses API has an mcp tool type with server_url, allowed_tools and require_approval fields. It supports streamable HTTP and the older HTTP/SSE transport. For an MCPifex endpoint, set server_url to the base gateway URL and send the key as the authorization value, which OpenAI does not store, so you send it on every request. The MCP and Connectors guide has the details.
What do you do when ChatGPT cannot connect?
- The option is missing. Check your plan and your admin's connector policy; availability varies.
- ChatGPT says it cannot reach the server. Run Test connection on the instance in the portal first. If that passes, re-copy the full URL and check for a missing character or a trailing space.
- Error 401 or "unauthorized". The key is wrong or was revoked. Create a new key, rebuild the URL and update the connector.
- Connected, but no tools. Compare with the enabled tools on the instance. Refresh the connector and start a new chat.
- A tool runs and returns an error. The problem is on the data side: a database login, a Google permission or a rate limit. The instance's call log shows the error text.
- Authentication loops. You probably chose OAuth. Edit the connector and set No Authentication, because the key is already in the URL.
If your own server is the problem, OpenAI suggests testing it with the MCP Inspector tool before blaming ChatGPT. More MCPifex fixes are in troubleshooting.
What should you know about safety?
A connector hands a model access to real systems, so start narrow. Enable read tools first and use a database login that can only read. OpenAI warns that an untrusted MCP server can put sensitive data into the model's context. Only add servers you control or trust. For the prompt-injection angle, read prompt injection through MCP tools.
Quick answers
OpenAI's help article lists Business, Enterprise and Edu workspaces on the web; some guides also report Plus and Pro. Check Settings in your own account. Workspace admins can also restrict custom connectors.
A custom connector is how ChatGPT reaches a remote MCP server. You give ChatGPT the server's HTTPS address, and it lists and calls that server's tools. Built-in connectors for apps like Gmail are a separate, OpenAI-managed feature.
Not directly. ChatGPT needs a public HTTPS address, and a local stdio server has none. Host the server somewhere reachable, use a tunnel, or use a hosted gateway such as MCPifex.
Treat the URL like a password and share it with nobody. Only a revocable MCPifex key goes in it, never database passwords. If it leaks, revoke the key in the portal and create a new one.
Yes. The Responses API has an mcp tool that takes a server_url, an allowed_tools list and approval settings. It is meant for developers building apps, while the connector steps above are for the ChatGPT web app.
Sources
Ready to connect?
Run hosted MCP servers behind one endpoint and choose exactly which tools your AI can use.