Connect ChatGPT and Codex

Copy your workspace URL from Settings → Connect to OpenAI. See Work from Your Own AI for how connecting works and what a connected assistant can do.

OpenAI's two paths are two products rather than two scopes: the ChatGPT app keeps the server in your account there, and the Codex CLI keeps it for your machine. Neither has a per-project form.

In the ChatGPT app

Click your name at the bottom left, then Settings → Plugins → Add → Add MCP Server.

  1. Set Type to Streamable HTTP rather than STDIO — STDIO is for a server running on your own machine.
  2. Name it xpt_<your-workspace> and paste the URL.
  3. Leave Bearer token env var and Headers empty. The connection is authorized in the browser, so there is no token to put there.
  4. The server appears under Servers on the MCPs tab. Saving sometimes leaves you on Plugins instead, so switch tabs if you do not see it. Press Authenticate beside it: ChatGPT opens XPT Tracker in your browser. Sign in if you are not already, confirm on the page that appears, and the connection is live.

One server covers Codex inside the app and ChatGPT itself — the latter only in the Work tab, not Chat. If the tools seem to be missing, check which tab you are in before checking anything else.

In the Codex CLI

One command registers the server and starts the authorization in the same breath, printing a link to open:

codex mcp add xpt-tracker --url https://mcp.xpt.ai/<your-workspace>/mcp

It adds for your machine rather than for one project, so to reach two workspaces add two servers under different names.

If you close the browser window before finishing, the server is left configured but unusable. codex mcp list shows it with an Auth column reading Not logged in; this starts the authorization again:

codex mcp login xpt-tracker