Skip to content
Short answer: To add an MCP server to ChatGPT on the web, open Plugins, select the plus button, choose Add custom MCP server, enter the server’s URL and sign-in method, accept the risk warning and create it as a plugin. Then install the plugin and type @ in a chat to use it. ChatGPT connects to remote servers only, so a server needs a URL.
MCP servers give an AI assistant tools for other apps, such as your project board or your deployments. If the idea is new, start with our guide on what MCP servers are. This post covers ChatGPT. Coding agents work differently, so we cover Codex and Claude in their own guides. To find servers worth adding, browse the full MCP servers directory.
ChatGPT MCP setup at a glance
| What to know | |
|---|---|
| Where it lives | Plugins in ChatGPT on the web |
| Starting point | The plus button, then Add custom MCP server |
| What you enter | The server URL and its sign-in method |
| Server type | Remote servers at a URL |
| Sign-in | OAuth is what OpenAI recommends |
| Result | The server becomes a plugin you install |
| Using it | Type @ in a chat and pick the plugin |
| Writes | ChatGPT asks you to confirm before any write action |
| Who can use it | Depends on your workspace permissions |
What you need first
- ChatGPT on the web. The custom server steps below start from the Plugins screen. Vercel’s docs say access to custom MCP servers depends on your workspace permissions and security restrictions, and OpenAI’s page does not list the plans that include it. If you do not see the option, ask whoever runs your workspace.
- The URL of a remote MCP server. Local servers that start with a command on your machine, the kind you add to a coding agent, do not work here.
- An account with the service behind the server, such as Vercel or Asana. The servers below are free to connect, but you need that account.
How to add a custom MCP server to ChatGPT
These are the steps in OpenAI’s documentation. Vercel’s docs walk through the same flow with a real server, so the example below uses it.
- In ChatGPT, open Plugins and use the plus button to select Add custom MCP server.
- Name the plugin and enter the server URL. For Vercel that is
https://mcp.vercel.com. - Choose the authentication method. Select OAuth if the server supports it.
- Read the connection warning, select I understand and want to continue, then select Create as a plugin.
- Install the plugin. Then type
@in a conversation and pick it from the list. - Test it with a prompt only that server can answer, in a normal chat or in deep research.
Some servers ask you to sign in to the service in your browser and approve access. Do that once. If the service later asks you to sign in again, repeat it.
Keep the first test small. A server that can only read is a gentler place to start than one that can create or delete, and you can always add a second server once the first behaves as you expect.
Which MCP servers to add first
Start with a remote server for an app you already use. These three list ChatGPT as a supported client on their Klyvoo pages and run at a URL:
- Vercel MCP: lists projects and deployments, reads logs and queries Web Analytics. The URL is
https://mcp.vercel.comand the sign-in is OAuth. Vercel says it is available on all plans. - Asana MCP Server: searches, creates and updates tasks, projects and portfolios. The URL is
https://mcp.asana.com/v2/mcpand the sign-in is OAuth. Asana’s page says the client app must not be blocked in your workspace. - Airtable MCP Server: searches and edits records in your bases. The URL is
https://mcp.airtable.com/mcp. Airtable lists OAuth or a personal access token as sign-in options.
Add one, test it, then add another. Every server adds tools for ChatGPT to choose from, and a short list keeps its choices sharper. Asana also offers an Asana app inside ChatGPT, and its help center has the steps. Airtable links a ChatGPT quickstart from its docs. When a vendor has its own ChatGPT page, follow it over a generic guide.
Read and write actions
A server can read your data, change it or both. OpenAI says ChatGPT currently asks for manual confirmation before any write action in a conversation. It also warns that a write can occur even if a server labels an action as read-only. So read each confirmation before you approve it.
Limits come from the service. Airtable, for instance, caps record creation per request and only lets you reach data your own permissions allow. Vercel says connecting grants the AI the same access as your Vercel user account. Treat the plugin as you, because to the service it is.
Business and Enterprise workspaces
In a workspace, an admin decides what people can do. OpenAI’s documentation says the Use plugins permission is on by default for Business and Enterprise workspaces, while the two MCP publishing permissions are off. Admins can also control which plugins and tools are available. If you cannot add a server, that is the first thing to ask about.
Asana has its own layer. Its docs say Enterprise+ admins can allow or block individual MCP clients in app management.
Check that it works
- Open a new chat and type
@to see whether your plugin is in the list. - Ask for something small and read-only, such as “list my five most recent deployments” for Vercel.
- Check the answer against the service itself, so you know the plugin sees the account you expect.
Common errors and fixes
- No Add custom MCP server option. Your workspace may restrict it. Ask an admin to check the plugin and MCP permissions.
- The sign-in fails. Remove the plugin, add it again and complete the sign-in from the start. Asana suggests logging out and back in, and checking your workspace permissions, if authentication fails.
- The plugin is missing from the @ list. It must be installed after you create it. Open Plugins and install it.
- The server rejects ChatGPT. Some vendors review which clients may connect. Vercel, for one, only allows clients it has approved and lists ChatGPT among them. Check the vendor’s page for the same list.
- A local server will not add. ChatGPT needs a URL. If the server only runs on your machine, use a coding agent such as Codex instead.
What MCP servers in ChatGPT cannot do
A server only exposes what its maker built, and only within what your account allows. OpenAI says custom MCP servers are not developed or verified by OpenAI, and any server receives whatever data ChatGPT puts in its queries, including things you shared earlier in the chat. A malicious or careless server can also be a path for prompt injection, where hidden instructions in the data it reads try to steer ChatGPT. OpenAI’s advice is to prefer official servers hosted by the service provider and to limit access to servers that hold sensitive data. For a closer look at how MCP differs from a plain API, read MCP vs API.
FAQ
How do I add an MCP server to ChatGPT?
Open Plugins in ChatGPT on the web, select the plus button and choose Add custom MCP server. Enter the server URL and sign-in method, accept the warning and create it as a plugin. Then install the plugin and use @ in a chat to call it.
Does ChatGPT support local MCP servers?
The custom server flow takes a server URL, so it is built for remote servers. A server that only runs as a command on your machine needs a client that can start it, such as a coding agent like Codex or Claude Code.
Which plans can use custom MCP servers in ChatGPT?
OpenAI’s MCP documentation does not name the plans, and Vercel says access depends on your workspace permissions and security restrictions. If the option is missing, ask your workspace admin. We left plan names out because we could not confirm them on OpenAI’s own page.
Is it safe to connect an MCP server to ChatGPT?
It is as safe as the server and the access you give it. OpenAI says custom servers are not verified by OpenAI and recommends official servers from the service provider. ChatGPT asks you to confirm write actions, so read each one before approving.
Can ChatGPT change data through an MCP server?
Yes, if the server offers write tools and your account may use them. ChatGPT asks for manual confirmation before a write action. Read-only servers or read-only sign-in scopes keep it to looking.
Are MCP servers free to use in ChatGPT?
Adding a server costs nothing, and the three above are free to connect. You still need an account with the service behind each one, and that service sets its own prices.
These steps come from OpenAI’s MCP documentation for ChatGPT and Vercel’s MCP docs, checked in October 2026. The ChatGPT screens change often, so follow OpenAI’s page if a label differs. For more servers to try, browse the Klyvoo MCP servers directory or see how ChatGPT Deep Research uses connected tools.