Back to blog

ChatGPT · MCP · AI Agents

Give ChatGPT a YouTube Transcript: Paste, Upload, Custom GPT or MCP

Four ways to get a video's words into ChatGPT: paste, file upload, a custom GPT with Actions, or a custom MCP server. What OpenAI's docs say today, what we measured on the transcript side, and what we could not test.

· 7 min read · YTAPI

ChatGPT can work with a YouTube video's transcript in four ways: you paste the text into the chat, you upload it as a file, a custom GPT fetches it through an Action, or a custom MCP server fetches it as a tool. Paste and upload work with any text. The other two make ChatGPT fetch transcripts itself, and as of November 15, 2026 OpenAI's own docs say custom GPTs are being replaced by plugins, so the MCP route is the one to build on.

We make YTAPI, a transcript API, so the examples use it. One limit up front: we did not drive the ChatGPT interface for this post. We have no ChatGPT account set up for testing, and OpenAI's Help Center blocked our automated requests. Everything about ChatGPT below comes from OpenAI's developer and Learn docs, read on November 15, 2026, and is marked as untested. What we ran ourselves is the transcript side: sizes, and a check of our MCP endpoint against OpenAI's checklist.

The four ways at a glance

WayWhat ChatGPT needsStatus in OpenAI's docs (Nov 15, 2026)Tested in ChatGPT by us
Paste the textNothingNot a documented feature, it is just a messageNo
Upload a fileA file with the textNot read from the docs (Help Center blocked us)No
Custom GPT with ActionsA GPT you createCustom GPTs are moving to plugins; actions do not transferNo
Custom MCP serverA public HTTPS MCP endpointDocumented under Plugins: Add custom MCP serverNo (endpoint checked from outside)

How big is a transcript?

Whether paste or upload works depends on length. We fetched three videos with the YTAPI REST API on November 15, 2026, as plain text:

VideoLengthCharactersWords
3Blue1Brown, "But what is a neural network?"18:4018,4303,357
Andrej Karpathy, "Let's build GPT"1:56:20108,99021,030
A Joe Rogan Experience episode2:51:54176,53934,528

Roughly 3,000 words for every 17 minutes of speech. A short video is a normal chat message. A podcast episode is a book chapter, which is where paste limits and file handling start to matter. We did not measure ChatGPT's own limits.

1. Paste the transcript

You get the text from somewhere (a transcript API, a download tool, or the transcript panel on YouTube), copy it and paste it after your question. Nothing to set up, works on any plan, and you can trim it to the part you care about.

The costs are manual: one copy-paste per video, no timestamps unless your source includes them, and long transcripts may not fit in one message. For one or two videos this is the right answer. For ten or more, use the fetching routes below.

With YTAPI you can get the text with one request (format=text returns plain text; segments adds timestamps):

curl -H "Authorization: Bearer $YTAPI_KEY" \
  "https://api.ytapi.dev/v1/transcripts?video_id=aircAruvnKk&format=text"

2. Upload a file

Save the transcript as a .txt or .md file and attach it to the chat. We could not read OpenAI's current file-upload page (it returned HTTP 403 to our requests), so we are not quoting size, token or per-day limits here. Check the file uploads page in the OpenAI Help Center for the limits on your plan.

Upload helps when the transcript is too long to paste comfortably, and when you want one file per video with the title in the file name. It still means fetching and saving each transcript yourself.

3. A custom GPT with Actions

This used to be the standard answer: create a GPT, import an OpenAPI file as an Action, add an API key, and the GPT calls the API itself.

OpenAI's guide Moving your custom GPT workflows to plugins (read November 15, 2026) says:

  • "We're transitioning custom GPTs to plugins."
  • At retirement, "custom GPTs stop running and leave the GPT directory." The page we read gave no retirement date.
  • Existing GPTs can keep being edited until retirement, and "custom actions need to be rebuilt separately". A custom action does not transfer to a plugin; OpenAI points to a supported app's actions or a custom MCP server as the replacement.
  • The page is written for ChatGPT Enterprise admins and creators. It says timelines for other plans may differ.

So building a new custom GPT around an Action is building on something OpenAI says is going away. If you already have one, it should keep working until retirement.

For reference, YTAPI publishes an OpenAPI 3.1.0 file at https://ytapi.dev/openapi.json with 15 operations, including getTranscriptByQuery. We checked that it downloads and parses. We did not import it into the GPT editor, so we cannot say how that editor treats it today.

4. A custom MCP server

MCP (Model Context Protocol) is the open protocol for giving an AI app tools. OpenAI's Connect and test your plugin page (read November 15, 2026) describes adding one in ChatGPT:

  1. Go to ChatGPT Plugins.
  2. Select the plus button, then Add custom MCP server.
  3. Enter a name and description.
  4. Under Connection, enter the server URL including the /mcp path.
  5. Configure authentication, read the risk warning, and select I understand and want to continue.
  6. Select Create as a plugin, then review the tools ChatGPT found.

The same page lists what the server needs: a public HTTPS endpoint that supports streamable HTTP (typically at /mcp), tool names, descriptions, schemas and annotations, and working authentication discovery. It also says "account and workspace policies apply to adding and using custom MCP servers." In an Enterprise workspace, OpenAI's other pages say an admin controls who may create plugins with custom MCP servers; that permission is off by default. We could not confirm from OpenAI's pages which individual plans (Free, Go, Plus, Pro) can add a custom server, and several third-party articles contradict each other, so check your own account.

What we checked on the YTAPI endpoint

On November 15, 2026 we ran OpenAI's checklist against https://api.ytapi.dev/mcp from outside ChatGPT:

  • An unauthenticated request returns 401 with a WWW-Authenticate header that points to /.well-known/oauth-protected-resource/mcp.
  • That metadata names the authorization server https://ytapi.dev/api/auth and the scopes youtube:read and offline_access.
  • The authorization server metadata lists a dynamic client registration endpoint, PKCE with S256, and the authorization_code and refresh_token grants. It does not advertise Client ID Metadata Documents.
  • With an API key, initialize and tools/list work. The server lists 10 tools, and every one carries readOnlyHint: true.
  • get_transcript on the 2:51 podcast above returned the first 40,134 characters, then a line telling the model to call again with offset=39986. Later pages are free for 30 minutes. A model that follows that line reads the whole transcript in about five calls.

This shows the endpoint is shaped the way OpenAI's page asks. It does not show that ChatGPT connects to it, signs in, or picks the right tool. We have not seen that happen, and we will update this post when we have.

The server has no write tools, so there is nothing in it that changes your YouTube account or your YTAPI data. Sign-in is OAuth, so there is no key to paste into the chat. The same server works in Claude, and that setup we did run: see YouTube transcripts in Claude.

What it costs

A transcript costs 1 credit through the REST API or MCP. Fetching the later pages of a long transcript is free for 30 minutes. Video details and caption languages are free, and requests that fail, such as a video without captions, cost nothing. New accounts get 200 free credits with no card. Credits don't expire; packs start at $9 for 2,000.

What none of these can do

  • No captions, no transcript. If the uploader turned captions off and YouTube has not generated any, there is nothing to fetch. That request fails and costs nothing.
  • Public videos only. Private, members-only and age-restricted videos fail.
  • Only the words. A transcript has no slides, code on screen or charts, unless the speaker says them.
  • Auto-captions mishear names and jargon. Treat an odd term in a quote with suspicion.
  • Long transcripts are a lot of text. The Joe Rogan episode above is about 35,000 words. Ask about a part of the video, or have the model read it in pages.

Which one to use

  • One or two videos, once: paste, or upload if the text is long.
  • Many videos, or a chat you will keep coming back to: a custom MCP server, if your plan lets you add one.
  • An existing custom GPT with an Action: it works until OpenAI retires custom GPTs. Plan the move.
  • A whole channel: do not do it through chat. See transcripts for a whole YouTube channel or playlist.

FAQ

We did not test what ChatGPT does with a link. The words spoken in a video are not in the page's HTML; they sit in a separate caption file that a tool has to request. That is why every route above hands ChatGPT the text or a tool that fetches it.

Does this work with Claude too?

Yes, and we tested that one. The setup is shorter because Claude's connector sign-in is built in: YouTube transcripts in Claude.

Can I use YTAPI's API in ChatGPT without MCP?

Only through a custom GPT Action while custom GPTs last. OpenAI's migration guide says actions do not carry over to plugins.

Sources

All read on November 15, 2026:

Create a free account and try the transcript API.