Add a custom MCP connector to ChatGPT (2026 guide)

ChatGPT has supported custom MCP connectors since September 2025 — but what you can actually do with one depends on your plan, and the requirements for the server side are stricter than Claude's. Here is the full picture before you build.

Updated July 2026

The plan limits, first — because they decide everything

Before writing a single prompt or line of code, check what your ChatGPT plan allows. As of mid-2026, OpenAI gates custom MCP connectors by workspace type — verify against OpenAI's current help docs before committing to a plan, since these limits have shifted before:

Plan	Custom MCP connectors	Practical meaning
Free	No custom connectors	Only built-in connectors.
Plus / Pro (individual)	Read/fetch-only, via Developer Mode	Your server's search and read tools work; tools that create, update or delete will not be usable.
Business / Enterprise / Edu	Full connectors, including write-capable tools	The complete MCP tool surface is available.

This is the most common “my connector is broken” report that is not a bug: a Plus user adds a server with a create_issue tool and the tool never fires. Read-only tools on the same server work fine. If write actions matter for your use case, you need a Business/Enterprise/Edu workspace — or a client without this restriction, like Claude Desktop (setup guide).

What ChatGPT requires from the server

A public HTTPS URL. Remote MCP server URLs must use HTTPS — no local STDIO servers, no plain HTTP, no localhost tunnels for anything durable.
Streamable HTTP transport. The current MCP remote transport; legacy HTTP+SSE-only servers are on borrowed time across all clients.
OAuth for user-scoped auth. ChatGPT walks the standard MCP OAuth flow. One sharp edge: if the authorization server issues tokens without offline_access-style refresh, ChatGPT can lose access when the token expires and users must reauthenticate.

If you generate and host your server on BuildMyMCPServer, all three are the default: every server deploys to a public HTTPS endpoint speaking Streamable HTTP, behind an OAuth 2.1 authorization server with PKCE and Dynamic Client Registration. There is nothing extra to configure for ChatGPT specifically.

Setup, step by step

Get your server URL — from your own deployment, from a prompt-generated server, or by forking a template.
In ChatGPT: Settings → Apps & Connectors. On individual plans, enable Developer Mode under advanced settings first — the “create connector” option is hidden without it.
Add a new connector: name it, paste the MCP endpoint URL, select OAuth as the auth method.
ChatGPT registers itself with the authorization server and opens the consent screen. Approve; the connector shows as connected.
In a conversation, enable the connector (via the tools/plus menu) and ask for something only your tool can answer. Name the tool explicitly on the first test.

Verifying it actually works

ChatGPT is more eager than Claude to answer from its own knowledge instead of calling a tool. To force a real call, ask for data the model cannot know — a record you created today, a value behind your API. Then check the server's dashboard: a live tool-call log with latency and status per call is the ground truth for whether the connector fired.

Common failure modes

Symptom	Cause	Fix
“Create connector” option missing	Developer Mode off, or Free plan	Enable Developer Mode (Plus/Pro) or upgrade the workspace.
Write tools never execute	Individual plan — read/fetch-only restriction	Business/Enterprise/Edu workspace, or use a write-capable client.
Connector disconnects after hours/days	No refresh token (offline_access missing)	Reconnect; if you control the AS, enable refresh token issuance.
“Unable to reach server”	URL is not public HTTPS, or wrong endpoint path	Use the full https://…/mcp endpoint; no localhost, no http.

Is ChatGPT the right first client?

If your tools are read-only — search, lookup, reporting — ChatGPT works well on any paid plan and the setup above takes minutes. If your tools write data, start with Claude Desktop or Cursor where the full tool surface works on individual plans, and add ChatGPT when a team workspace exists. The server is the same either way; only the client config differs. Compare hosting options in our platform comparison, or check pricing — the free tier is enough to test a connector end to end.

Skip the boilerplate — describe your tool, get a hosted MCP server.

BuildMyMCPServer generates the TypeScript server, wraps it in OAuth 2.1 and deploys it to a public Streamable HTTP URL for Claude, Cursor and ChatGPT. Free tier, source export, no lock-in.

Start building →