Cursor MCP server: video transcripts inside the editor

Drop one JSON block into .cursor/mcp.json and Cursor's agent reads any video or podcast link you paste into chat.

Get an API key, 100 free credits/mo Set it up

Set it up in Cursor

Cursor reads MCP servers from .cursor/mcp.json, project-local, or ~/.cursor/mcp.json to have it in every project. Create the file (or use Settings → Tools & MCP → Add) with the mcp-remote bridge:

.cursor/mcp.json

{
  "mcpServers": {
    "transcriptfetch": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote", "https://transcriptfetch.com/mcp",
        "--header", "Authorization: Bearer tf_live_YOUR_KEY"
      ]
    }
  }
}

Cursor lists the new tools under Settings → Tools & MCP once the file saves; toggle them on and the agent can call them. Keys come from your dashboard (100 free credits a month).

What it can do

In chat or agent mode, Cursor can fetch the transcript of any video URL that appears in your conversation or your task, search YouTube, and list a channel or playlist. The editor-shaped win: paste a tutorial video next to the file you are working on and ask Cursor to apply what the video does to your code, no tab-switching, no manual transcription.

In Cursor chat

You: This video explains the migration we need. Apply the
     same steps to our schema: youtube.com/watch?v=...

Cursor: [get_transcript] The video migrates in three steps...
        Updating db/schema.ts to match:

Full MCP server overview →Configuration reference →Using Claude? →

Project keys and shared configs

The one decision Cursor forces is where the config lives. A project-local .cursor/mcp.json travels with the repo, which is convenient and dangerous in equal measure: commit it with a real key inside and the key is public history. Keep secrets in the global ~/.cursor/mcp.json, or commit the project file with a placeholder and let each teammate paste their own key.

Tool use in Cursor is agent-driven: you do not invoke get_transcript by name, you paste a link and ask for what you want, and the agent decides to call the tool. If it narrates instead of calling, say "use the transcriptfetch tool", once per session tends to be enough.

Transcripts land in the model's context like any other tool result, so an hour-long video is a big paste. For long material, ask for a summary or a specific section first; the timestamped segments make "what do they say around minute 40" a real question.

Frequently asked questions

Project config or global config?

Global (~/.cursor/mcp.json) for personal use with a real key. Project (.cursor/mcp.json) when a team should share the setup, with the key left as a placeholder so it never lands in git history.

Why doesn't Cursor call the tool?

Check the server is toggled on under Settings → Tools & MCP, then ask explicitly: 'use the transcriptfetch tool on this link'. Once the agent has used it in a session it reaches for it on its own.

Which platforms can it fetch?

YouTube, TikTok, Instagram, podcasts (Spotify, Apple, or RSS) and direct media file URLs. Search and channel/playlist listing are YouTube-only.

What does it cost?

100 free credits a month per account, one credit per successful transcript, failures never billed. Paid plans start at $5 a month for 500 credits.

Video sources, editor context

One JSON block and Cursor reads what videos say. 100 free credits a month to start.

Get started free