How to connect YouTube transcripts to Claude and ChatGPT

Wouldn't it be nice to paste a YouTube link into Claude or ChatGPT and just ask about it? Out of the box, these assistants cannot watch videos or read captions. The fix is the Model Context Protocol, or MCP, an open standard that lets an AI client call external tools (see the MCP server setup). This tutorial shows how to connect YouTube transcripts to Claude, Cursor, and ChatGPT in a few minutes, so your assistant can fetch, search, and summarize video content inline.

What MCP gives you

MCP is a common language between AI clients and tool servers. Once you connect the TranscriptFetch MCP server, your assistant gains four tools it can call on its own:

• get_transcript returns the full transcript for a single video.
• search_videos finds videos by keyword.
• list_channel_videos lists recent uploads for a channel.
• list_playlist_videos lists the videos in a playlist.

You do not write code. You ask in plain language, and the assistant decides which tool to call.

Step 1: get an API key

Create a key in your TranscriptFetch dashboard. It looks like tf_live_.... You will pass it to the MCP server as a bearer token so your usage and credits are tracked to your account. Keep the key private, the same way you would treat any password.

Step 2: connect Claude Code

Claude Code has a single command for adding an HTTP MCP server. Pass your key as an authorization header.

claude mcp add --transport http transcriptfetch https://transcriptfetch.com/mcp \
  --header "Authorization: Bearer tf_live_YOUR_KEY"

That is it. Start a conversation and the four tools are available. Claude will call them automatically when your question needs video data.

Step 3: connect Cursor or VS Code

Both editors read an MCP config file. Add a server entry with the URL and your authorization header.

For Cursor, edit ~/.cursor/mcp.json. For VS Code, edit .vscode/mcp.json in your project.

{
  "mcpServers": {
    "transcriptfetch": {
      "url": "https://transcriptfetch.com/mcp",
      "headers": {
        "Authorization": "Bearer tf_live_YOUR_KEY"
      }
    }
  }
}

Reload the editor and the tools appear in the assistant.

Step 4: connect ChatGPT and other clients

Any client that speaks the Streamable HTTP transport works the same way. Provide the endpoint URL and an authorization header.

{
  "url": "https://transcriptfetch.com/mcp",
  "headers": {
    "Authorization": "Bearer tf_live_YOUR_KEY"
  }
}

In ChatGPT, add this as a connector or custom MCP server in the settings that support it. The exact menu changes over time, but the two things you always supply are the URL and the bearer header above.

Step 5: try asking

Once connected, talk to your assistant normally. Useful prompts to start with:

Summarize the key points from this video: https://youtu.be/dQw4w9WgXcQ

Find videos about vector databases and summarize the top three.

Pull the latest 10 uploads from @channelhandle and list the topics they cover.

Read this playlist and tell me which video best explains embeddings.

The assistant calls the right tool, reads the transcript, and answers using the actual content of the videos rather than guessing.

How auth and billing work

• The assistant sends your bearer key with every request, so calls are tied to your account.
• A connection without a valid key is rejected during the handshake, so a bad key fails fast.
• Each successful fetch costs one credit. Videos with no captions return an empty result and are free.
• Calls are rate limited per key, exactly like the REST API.

Troubleshooting

• The server will not connect. Double check that you included the Authorization header. The endpoint requires a valid key even to list the tools.
• Tools do not appear. Reload the client after editing the config, and confirm the URL is https://transcriptfetch.com/mcp with no trailing path.
• A video returns nothing. It probably has no captions. Try another video to confirm the connection itself is working.

Wrapping up

MCP turns YouTube into something your AI assistant can actually use. Add the server once with your key, then ask questions in plain language and let the assistant fetch the transcripts for you. For the full tool reference and client specific notes, see the MCP documentation.