How to Connect ChatGPT to a YouTube MCP Server

ChatGPT connects to Prepublish by adding it as a developer-mode app pointed at https://mcp.prepublish.ai. Two things must happen first: developer mode has to be switched on for the account, and then an app has to be created from ChatGPT Plugins, scanned, and added to a conversation from the composer's Developer mode menu.

Prepublish audits a script you have written but not recorded. ChatGPT is a remote-only MCP client, so the endpoint is the whole configuration and there is no local process to install. The shared reference, including the six tools and the published limits, is the Claude Code hub guide. This page covers only what is specific to ChatGPT.

OpenAI documents the flow in two places, and they name slightly different menus. Both were read on 2026-09-21, and the difference is worth knowing because it saves you hunting.

• ChatGPT developer mode says to open Settings → Security and login in ChatGPT and turn on Developer mode, and lists eligibility as Pro, Plus, Business, Enterprise and Education accounts on the web.
• Developer mode and MCP apps in ChatGPT describes the administrator path. On a Business plan an admin enables it from Workspace Settings → Permissions & Roles → Connected Data, where the toggle is labelled Developer mode or Create custom MCP connectors. On Enterprise and Edu plans an admin can enable it from Settings → Apps → Advanced Settings, or grant access through RBAC first.

Then create the app, again from the developer mode guide:

1. Open ChatGPT Plugins and select the plus button.
2. Create a developer-mode app for a remote MCP server and provide the endpoint, https://mcp.prepublish.ai, with the metadata it asks for.
3. Choose the authentication mechanism. Prepublish is anonymous, so choose no authentication. An optional pp_live_ API key is a header, not an OAuth flow, so it belongs in a client that can set headers rather than here.
4. Click Scan Tools and wait for the scan to finish.
5. Click Create. The app appears under Drafts in app settings, and its tools become available in the composer's Developer mode tool.

Both OpenAI pages agree that MCP apps run on the web. The help article answers the mobile question directly: apps are not available on mobile.

The supported transports are SSE and streaming HTTP, per the developer mode guide. Prepublish serves Streamable HTTP, which is the transport that guide calls streaming HTTP, so no configuration beyond the URL is required.

Every other client on this site has two routes, the hosted URL and the npx -y prepublish-mcp bridge. ChatGPT has one, because it is a remote-only MCP client.

The help article read on 2026-09-21 states that ChatGPT cannot connect to a local MCP server directly, and that a server on a private network, on premises, or on a developer machine needs Secure MCP Tunnel instead. The bridge's job is the opposite: it starts a process on your machine and re-serves the hosted server over stdio, which is what a desktop client needs when it can only launch a command. ChatGPT never launches a command, so the bridge has nothing to attach to.

The practical consequence is that the decision the hub guide describes does not apply here. Paste the endpoint. https://mcp.prepublish.ai is canonical, and https://prepublish.ai/mcp and https://mcp.prepublish.ai/mcp serve the same handler, so an entry copied from older documentation keeps working.

Four failures belong to the client rather than to the server.

The tool scan did not finish. Without a completed scan there are no tools to select, and the app stays a draft. Re-run the scan from the app settings page, and if the server is OAuth-protected, finish the authorization prompt first. Prepublish is not OAuth-protected, so a scan against this endpoint has no authorization step in the middle.

The tool list is frozen. The help article read on 2026-09-21 states that after an admin first approves an MCP app for a workspace, ChatGPT uses a frozen snapshot of the available tools and inputs, and that changes made later by the developer are not applied until an admin reviews and re-publishes an update. It also states that if the live app no longer matches the snapshot, tool calls can error, and that backward-compatible changes such as a new optional parameter may keep working.

The account tier limits what tools can do. The help article says full MCP support including write actions is in beta on Business, Enterprise and Edu plans, and that Pro users can connect MCPs with read and fetch permissions in developer mode. Prepublish's two read-only tools, script_runtime and get_audit, declare that property. The four AI-backed tools ask for confirmation rather than running silently, which is normal for a tool that is not marked read-only.

An administrator has to enable it on a workspace plan. On Business plans the help article says only admins and owners can enable developer mode, that each admin enables it for themselves, and that admins cannot enable it for individual members.

When the problem is on the server side, the message says what it is. The strings come from the endpoint's own code.

Because ChatGPT makes its requests from OpenAI infrastructure rather than from your browser, an origin check matters: an absent Origin is allowed, and only a present, unlisted one is refused. The deployment lists its own origins and can add another through the MCP_ALLOWED_ORIGINS environment variable, which is the documented fix when a hosted connector presents an origin nobody has seen.

Select the app in the conversation first. Open the plus menu, choose Developer mode, and pick Prepublish. Then name the tool you want, because ChatGPT chooses between tools from their descriptions and a direct instruction removes the guesswork.

Use the script_runtime tool. How many words do I need for a twelve minute video, and how long would a 2,400 word script run?

The response carries speaking_rates_words_per_minute with fast, median and slow, a source string, a caveat string, word_count, runtime_minutes and word_budget, the last two with the same three keys. The rates are the measured ones in prepublish-fe/lib/seo/wpm-data.ts. Apply the arithmetic and a 2,400 word script is 15 minutes at 160 words per minute, 13.3 at 181 and 11.9 at 201, while a twelve minute target needs 1,920 words at the slow rate, 2,172 at the median and 2,412 at the fast rate. Nothing is billed for that call, and no email is asked for.

Then ask for the check the product is built around:

Use the audit_hook tool on this opening. Tell me which sentence loses attention first: "Most creators lose half their viewers in the first fifteen seconds, and the reason is rarely the editing."

audit_hook takes hook_text, an optional niche and an optional email, and returns a hook object with evaluation_id, hook_text, niche, overall_score, grade, one_sentence_verdict, sentences, top_issues, rewrites, model_used and created_at. Each entry in sentences carries its own attention read. This tool is AI-backed, so it counts against the allowance, and ChatGPT asks you to confirm the call because the tool is not marked read-only.

Every AI-backed payload repeats the same limit, and the model relays it: "This is a text-only check of an unrecorded script. It maps relative attention risk inside the draft. It does not measure or predict published YouTube retention, and it cannot account for delivery, editing, thumbnail, topic or distribution."

The check reads text. It does not watch a video, fetch a transcript or read your channel, and no result is a forecast of audience behaviour.

The free tier is published in prepublish-fe/lib/mcp/limits.ts, and the live numbers for the current day are in the server card at https://mcp.prepublish.ai.

A full audit is released by email, so audit_script asks for an address and the backend rejects a request without one. The narrower checks do not ask. On a workspace plan the frozen tool snapshot is a second limit that has nothing to do with quotas: a tool definition that changed after approval can error until an admin refreshes the app.