Give Cursor YouTube video access
A block in .cursor/mcp.json gives Cursor 22 tools for reading YouTube videos — plus one credential mistake worth avoiding first.
The short answer
That opens Cursor and asks to install the server. Approve it, then click Needs login next to it in Settings → MCP and sign in once. If nothing happens when you click, Cursor is not installed on this computer — the file below does the same thing by hand.
Or create .cursor/mcp.json in your project with the URL and no credential at all:
{
"mcpServers": {
"vidwords": {
"url": "https://vidwords.com/mcp"
}
}
}
Cursor picks it up without a restart in most cases. The MCP section of Settings shows the server as Needs login — click that once, approve in the browser, and Cursor holds a credential it refreshes by itself. Use ~/.cursor/mcp.json instead to make it available in every project.
Prefer this to a pasted token. The file above carries no secret, so the committed-credential problem described below cannot arise.
A static token instead
If your Cursor build does not offer the sign-in step, add the header back — and read the warning two sections down before you commit anything:
"headers": { "Authorization": "Basic YOUR_API_TOKEN" }
Get a token first
Only the static-token route above needs this; signing in issues its own credential. Create a free account and create a key on the API & MCP page (its value is shown once). Two things to know before the first call:
- Verify your email. Until you click the verification link every call returns
403with{"error":"email_unverified"}— the most common first-call failure on a new account. - The scheme is
Basic, notBearer, and the token goes in as-is. You are not base64-encoding auser:passpair.
Do not commit the project file
This is the one thing worth getting right before anything else. .cursor/mcp.json sits inside the repository and gets committed by default — and the header holds a live API token. Committing it hands a working credential to everyone with repository access, and to the entire internet if the repo is public.
Either put the server in ~/.cursor/mcp.json, which is outside the repo and covers every project anyway, or add .cursor/mcp.json to .gitignore before your first commit. If a token has already been pushed, rotate it — deleting the file does not remove it from history.
Using it in Agent mode
MCP tools are available to Cursor's agent, which means the model decides when to call them. Two habits make that reliable:
- Give it the URL. "Summarise the React Conf keynote" invites an answer from memory; pasting the link makes the tool the obvious route.
- Ask for timestamps. It steers the model toward
search_transcriptrather thanget_transcript, and you get citations you can click instead of a summary you have to take on faith.
The workflow this unlocks in an editor is narrow but genuinely underserved: a great deal of framework knowledge is published as conference talks and screen recordings and never written down. Being able to ask "what did they say the migration path was, with timestamps" without leaving the editor removes a context switch that otherwise costs half an hour.
Reading what is on the screen
Screen-recorded tutorials are the case where captions fail completely. The presenter says "as you can see here" and the actual content — the config, the terminal output, the diagram — exists only in the picture. analyze_video reads frames as well as speech, and ask_video answers against that analysis with every citation checked against a recorded frame; anything that cannot be matched is dropped rather than guessed at.
The 22 tools
| Tool | What it does | Cost |
|---|---|---|
search_transcript | Find where a video — or a list of up to 25 — discusses something; returns timestamps and deep links. | 1 Cloud Request per new video |
get_transcript | Full text for up to 25 videos, a span between timecodes, or a page at a time with maxChars. | 1 Cloud Request per new video |
translate_transcript | One video’s transcript translated into any language by AI, line for line, every timestamp kept — for a language the video has no captions in, or an audio transcription. estimateOnly quotes the price first; reading it again is free. | 2 AI Units per started minute of video, only when it succeeds |
search_library · list_library | Search and list the transcripts you have already saved. | Free |
list_languages | A video's caption tracks, and which ones transcribe the speech. | Free |
list_channel_videos | Recent uploads for a channel. | 1 Cloud Request · Starter and up |
list_watchlists · watchlist_activity | Radar monitoring — channels you track and what they published. | Free |
create_watchlist · add_watchlist_channel · check_watchlist · preview_watchlist_filters | Set up Radar: create a watchlist, add channels, check them for new uploads now, and test title filters before saving them. | Free |
update_watchlist · set_radar_paused | Rename, pause or filter a watchlist, or pause all of Radar. Turning on automatic analysis spends AI Units at Watch rates, so an assistant should only do it when you ask. | Free; automatic analysis is billed |
remove_watchlist_channel · delete_watchlist | Stop monitoring a channel, or delete a watchlist and the videos it recorded. | Free |
account | Plan and remaining Cloud/AI balances. | Free |
analyze_video | Frame-level analysis — slides, charts, demos, on-screen text. estimateOnly quotes the price first. | 1 Cloud Request, then AI Units per minute: Standard analysis 4, Deep 30, Quick 2.8 |
list_analyses | Your AI Watch analyses, newest first, so one can be reopened. | Free |
get_analysis | Read a finished analysis; waitSeconds waits for it. | Free |
ask_video | Ask against a finished analysis; citations verified or dropped. | 6 AI Units per question |
Re-reading a video the account already has is free — another time range, another page, or a search after a fetch — and every metered result reports what it charged.
If it does not connect
- Server shows as failed in Settings. Check the header is
Basic, notBearer, and that the JSON parses. - 403 on every call. Unverified email — click the link from signup.
- Tools listed but never used. Agent mode has to be active, and the request needs to make a video the obvious subject. Paste the URL.
- 429 during a wide search. The limit is per minute by plan — 60 on Free, 200 on Starter, 500 on Pro, 1,000 on Team — and a stateless server re-handshakes on every call, so a fan-out across many videos reaches it quickly.