How to Add an MCP Server to ChatGPT, Claude Desktop, and Cursor

To add an MCP server to ChatGPT, Claude Desktop, or Cursor you need the same three things and three different screens. Here are the exact menu names as the vendors document them as of September 2026; every plan or eligibility fact is the vendor's, never asserted here.

What you need before you start

Three things: a server the client can reach, an account the vendor lets in (the table below quotes the vendors on plans), and a way to authenticate.

A remote MCP server, reachable over HTTPS

An MCP server advertises tools and runs them on request, as a local process (stdio) or as a remote MCP server behind an HTTPS URL. ChatGPT's web app connects only to remote servers, per OpenAI's developer mode article; a laptop server needs what the article calls a Secure MCP Tunnel. Claude's custom connectors are remote by definition, per Anthropic's help page. Cursor takes both. So the URL that works everywhere is a public https:// endpoint, usually ending in /mcp; whether it should be an MCP server at all is a separate decision.

MCP authentication: OAuth or a token

Remote servers need to know who is calling: with OAuth the client opens your browser, you approve once, and it keeps a refreshable token; with a static credential you paste a bearer token into the client's config. All three clients support the browser flow, so try it first; the server decides which it wants.

How to connect an MCP server to ChatGPT

ChatGPT MCP client support comes in two surfaces that look nothing alike: the desktop app keeps servers under Plugins, the web app under Apps behind developer mode. OpenAI's article documents the web surface; the desktop steps match the app's own screens, which this ChatGPT setup walkthrough shows filled in.

The ChatGPT MCP client on desktop: Plugins, then MCPs

1. Open Settings, then Plugins under Integrations, and switch to the MCPs tab.
2. Choose Add MCP server and name it.
3. Pick Streamable HTTP, not STDIO, which is for a process ChatGPT launches on your machine.
4. Paste the HTTPS URL, leaving the bearer token env var and both header tables empty for OAuth.
5. Save, choose Authenticate on the new row, and approve the consent page in your browser.

If there is no MCPs tab, whether your account shows one is documented by OpenAI, not here. One distinction: OpenAI's "web only" answers the mobile question; the desktop app's Plugins, MCPs screen is a separate surface its article does not cover, so read neither as covering the other.

The ChatGPT MCP client on the web: Apps and developer mode

Everything here is from OpenAI's article as of September 2026.

1. Turn on developer mode: Settings, then Apps, then Advanced settings (Pro accounts, Enterprise and Edu members); on a Business workspace only admins and owners can, from Workspace settings, Apps, Create.
2. Open Apps and choose Create.
3. Enter the endpoint and its metadata, then pick the auth method the server expects.
4. Choose Scan Tools, completing the OAuth prompt if there is one.
5. Choose Create. The app appears as a Draft with a Dev label.

Two more facts from the article: a server no longer needs search and fetch tools, and an approved app runs on a frozen snapshot of its tool list until an admin republishes.

How to add a Claude Desktop custom connector

Anthropic calls a remote MCP server a custom connector; one screen serves Claude Desktop and claude.ai, and the Claude Desktop page in these docs shows it filled in. Steps per the MCP project's remote-server guide and Anthropic's help article, as of September 2026.

1. Press Ctrl+Comma, or open the menu icon, then File, then Settings.
2. Choose Connectors in the sidebar, then Add at the top right (Anthropic words it Customize, Connectors, then the plus button).
3. Choose Add custom connector and paste the remote server URL. Advanced settings takes an OAuth Client ID and Secret, both optional.
4. Choose Add and complete the authentication prompt in your browser.
5. Set the connector's tool permissions: per tool, what Claude may call without asking.

Anthropic adds three warnings: connect only servers you trust, expect prompt injection through what a tool returns, and disable write tools during Research. On Team and Enterprise only Owners add custom connectors (Organization settings, Connectors, Add, Custom, Web), and each member then clicks Connect to authenticate as themselves.

Cursor mcp.json: add a remote server in one file

Cursor keeps MCP servers in a file, not a form. Per Cursor's MCP docs as of September 2026, that file is .cursor/mcp.json in a project (that project only) or ~/.cursor/mcp.json in your home directory (every project). A remote server is an entry under mcpServers with a url:

{
  "mcpServers": {
    "example-tools": {
      "url": "https://example.com/mcp"
    }
  }
}

Save it and Cursor picks it up; a server that uses OAuth completes sign-in through Cursor's OAuth callback, so try that first. For a static credential, add "headers": { "Authorization": "Bearer <your token>" } to the entry and keep the real value in the global file, not a shared repository. The docs also describe an optional auth block (CLIENT_ID, CLIENT_SECRET, scopes) for a fixed OAuth client.

Which transport and auth do these clients expect?

Why does ChatGPT say it cannot use my MCP server?

Five causes cover most reports, each from OpenAI's article or the form.

• The URL is local. The web app connects to remote servers only; OpenAI's answer for a laptop server is its Secure MCP Tunnel.
• Developer mode is off, or your account lacks it. Apps, Create needs it, and OpenAI ties it to account type and role; on Business only an admin or owner can turn it on.
• You are on mobile. The article says MCP apps are web only.
• The tool list is stale. An approved app runs on a frozen snapshot of its tools until an admin republishes.
• The token expired. OpenAI tells server authors to issue refresh tokens by supporting offline_access; a server that does not forces re-authentication, which reads like "cannot use" from the chat side.

Two more are not ChatGPT's doing: the server is down, or the desktop form was given STDIO instead of Streamable HTTP.

What can a connected MCP server do inside a chat?

Read, and with confirmation, write. On ChatGPT web, OpenAI's article says a write action prompts for confirmation, agent mode will not use custom apps, and deep research uses them read-only. Claude sets permissions per tool and recommends disabling write tools during Research. Cursor asks for approval before using an MCP tool by default, per its MCP docs as of September 2026, and its Auto-review run mode runs allowlisted tools immediately; either way, expose fewer write tools, or none, to a coding agent.

The vendors' warnings apply everywhere: connect only servers you trust, and treat every tool result as untrusted input, because a server returns text the model reads as if you had typed it. Prompt injection through a tool result is the standard attack on a connected agent; our post on prompt injection in agents has the containment checklist. Keep write tools off until the read tools have behaved, and never paste a real credential into a shared file.

Behind the URL can sit a thin wrapper over a database or a whole hosted service; a rented Expert on Askpert is one example of the second kind, and the client pages under Connect your agent walk the same screens.

MCP client setup FAQ