VidWords YouTube MCP Server

A hosted Model Context Protocol server that lets an AI agent read YouTube videos — and cite the exact second it got the answer from.

A language model cannot watch a video. Point it at this endpoint and it gains thirteen tools for searching transcripts, searching the videos you have already saved, reading a video's frames — slides, charts, demos, on-screen text — and answering questions with citations that are verified before you see them.

No integration code. No scraping. No proxy pool.

https://vidwords.com/mcp

That URL is the whole configuration. Clients sign in over OAuth — no key to find or paste — and headless ones can send Authorization: Basic <your-api-token> instead.

Remote-only and hosted — there is nothing to install or self-host. This repository is the public manifest, configuration reference and issue tracker for that endpoint.

Quick start

Most clients need no token at all. The server speaks OAuth, so the client registers itself, sends you to VidWords to sign in, and stores a credential it refreshes on its own. You can create the account during that sign-in step. The free plan includes 25 Cloud Requests and 200 AI Units a month, so you can wire this up and use it before paying anything.

Claude (claude.ai and Claude Desktop)

1. Customize → Connectors, then + Add → Add custom connector.
2. URL: https://vidwords.com/mcp
3. Under the OAuth client options choose Register automatically. VidWords registers each connector itself (dynamic client registration), so Claude's default "published identity" option will not connect.
4. Connect, sign in to VidWords, and approve.

On a Team or Enterprise plan an owner adds it first, under Organization settings → Connectors → Add → Custom → Web; members then connect from Customize → Connectors.

ChatGPT

1. On chatgpt.com, open Plugins, select +, then Add custom MCP server.
2. Name: VidWords · Server URL: https://vidwords.com/mcp · Authentication: OAuth.
3. Accept the risk warning and select Create as a plugin, then sign in to VidWords and approve.
4. Install the new plugin, then type @ in a chat and pick VidWords.

Do this on the web; on a workspace account an admin may need to allow custom MCP servers.

Either way, the host registers itself, sends you to VidWords to sign in, and shows a consent screen naming exactly what it is asking for. Registration alone grants nothing — access begins only when a signed-in person clicks Approve, and live connections can be revoked from the API & MCP page with immediate effect.

Claude Desktop from a config file

claude_desktop_config.json only runs local (stdio) servers — a remote url entry there is not supported (and has been reported to wipe the file's mcpServers section). To declare the server in the file anyway, bridge it with mcp-remote, which runs the same sign-in flow in your browser — still nothing to paste:

{
  "mcpServers": {
    "vidwords": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://vidwords.com/mcp"]
    }
  }
}

Quit and reopen Claude Desktop after editing the file.

Claude Code

claude mcp add --transport http vidwords https://vidwords.com/mcp

Then type /mcp in a session and choose Authenticate.

Cursor — one click, or .cursor/mcp.json

Use the Add to Cursor button above, or:

{
  "mcpServers": {
    "vidwords": {
      "url": "https://vidwords.com/mcp"
    }
  }
}

Cursor shows the server as Needs login — click that once and it runs the OAuth flow in your browser. Because this file carries no secret, it is safe to commit, which the header form below is not.

VS Code

Use the Add to VS Code button above, or put this in .vscode/mcp.json:

{
  "servers": {
    "vidwords": { "type": "http", "url": "https://vidwords.com/mcp" }
  }
}

VS Code asks you to sign in the first time the server starts.

Codex CLI

codex mcp add vidwords --url https://vidwords.com/mcp
codex mcp login vidwords

A static token instead

For CI, a container, or a client with no OAuth support, authenticate with a header. Create an account at vidwords.com/register, verify your email, then create a key on the API & MCP page. A key's value is shown once, when it is created.

Claude Code

claude mcp add --transport http vidwords https://vidwords.com/mcp \
  --header "Authorization: Basic YOUR_API_TOKEN"

Claude Desktop — claude_desktop_config.json

The same mcp-remote bridge as above, with the key instead of the sign-in. Keep the key in env: Claude Desktop on Windows (and Cursor, and the Codex CLI) do not escape spaces inside args, so Basic YOUR_API_TOKEN written there splits in two — the mcp-remote README's own workaround:

{
  "mcpServers": {
    "vidwords": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://vidwords.com/mcp",
               "--header", "Authorization:${VIDWORDS_MCP_AUTH}"],
      "env": { "VIDWORDS_MCP_AUTH": "Basic YOUR_API_TOKEN" }
    }
  }
}

Cursor — .cursor/mcp.json

{
  "mcpServers": {
    "vidwords": {
      "url": "https://vidwords.com/mcp",
      "headers": { "Authorization": "Basic YOUR_API_TOKEN" }
    }
  }
}

Keep this out of version control, or use ~/.cursor/mcp.json instead — the header holds a live credential.

Codex CLI — ~/.codex/config.toml

[mcp_servers.vidwords]
url = "https://vidwords.com/mcp"
env_http_headers = { "Authorization" = "VIDWORDS_MCP_AUTH" }

export VIDWORDS_MCP_AUTH="Basic YOUR_API_TOKEN"

Do not use bearer_token_env_var. It is the obvious-looking field, but it sends Authorization: Bearer <value> and this server authenticates with Basic.

Clients without custom-header support, and Docker

This repository also ships a small stdio proxy (src/index.js) that speaks MCP on stdin/stdout and forwards tool calls to the hosted endpoint. Use it when your client cannot send a custom HTTP header, or when you want the server in a container:

{
  "mcpServers": {
    "vidwords": {
      "command": "npx",
      "args": ["-y", "github:haljishi/vidwords-mcp"],
      "env": { "VIDWORDS_API_TOKEN": "YOUR_API_TOKEN" }
    }
  }
}

Run straight from this repository — the proxy is not published to npm, so a bare npx @vidwords/mcp will not resolve.

docker build -t vidwords-mcp .
docker run --rm -i -e VIDWORDS_API_TOKEN=YOUR_API_TOKEN vidwords-mcp

The tool schemas are declared inline in the proxy, so initialize and tools/list answer without any credentials. Resource discovery is also local; the upstream is contacted when a tool, prompt, or UI resource is requested. The proxy also exposes the analysis and account MCP Apps resources. Compatible hosts can render interactive cards; the analysis card includes Show full analysis. UI resource reads and tool calls use the same authenticated upstream, and structured results pass through unchanged.

A call without VIDWORDS_API_TOKEN returns a readable error rather than failing the handshake. VIDWORDS_MCP_URL overrides the endpoint if you are pointing at a non-production instance.

The generic mcp-remote bridge works too:

{
  "mcpServers": {
    "vidwords": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://vidwords.com/mcp",
               "--header", "Authorization:${VIDWORDS_MCP_AUTH}"],
      "env": { "VIDWORDS_MCP_AUTH": "Basic YOUR_API_TOKEN" }
    }
  }
}

Ready-made config files live in examples/.

The thirteen tools

Two balances pay for them: a Cloud Request fetches one video's transcript, and AI Units pay for reading frames and for transcribing audio — a fresh frame analysis spends one of each kind: 1 Cloud Request for its transcript, plus AI Units per minute of video. A video the account has already fetched is in its Library, and reading it again is free — another time range, another page, a search after a fetch. Every metered result reports what it charged.

Check the Library first

search_library and list_library read what the account already has, for nothing. An agent that searches the Library before fetching a video again answers "what did that interview I watched last week say about pricing" without spending anything, and without the round trip.

Prefer search_transcript over get_transcript

Both cost one Cloud Request per new video, so there is no billing reason to choose. The reason is context. Ask "what did this two-hour interview say about pricing?" and get_transcript returns roughly 20,000 words, of which perhaps 300 are about pricing — those 300 now compete for attention with 19,700 that are not, and the answer gets worse, slower and more expensive to generate.

search_transcript returns only the matching stretches, each with a deep link. Reach for get_transcript when you genuinely want the whole text: an export, a diff, a corpus.

Ask for a span, not a whole video

Both transcript tools take optional from and to timecodes — seconds (615), m:ss (10:20) or h:mm:ss (1:02:13):

{ "videos": ["dQw4w9WgXcQ"], "from": "10:20", "to": "11:00" }

These are the same formats the tools print back, so a timestamp out of one answer can be pasted straight into the next question. A timecode that cannot be parsed is refused before anything is fetched, so a typo costs nothing — it never silently widens to the whole video.

Page a long transcript instead of guessing ranges

get_transcript takes maxChars. A result cut short says truncated: true and carries nextFrom, the second to continue from; pass it back as from. The cut always lands between captions, and every page after the first is a re-read of a video already in the Library — free.

{ "videos": ["dQw4w9WgXcQ"], "maxChars": 20000 }

Which language you got, and why

A caption track's language is not necessarily the language spoken. An auto-generated track transcribes the speech; an uploaded one may be a translation — some videos carry a dozen community translations and no track in the language actually spoken. So every transcript says why it is the track it is:

• selectionReason: requested, spoken_language (it matches the speech), caption_fallback (no track is known to match the speech — English, if present, was picked for readability), translated or audio_transcription.
• spokenLanguage: the language being spoken when it is known, otherwise null — not a guess.
• lang asks for a track. If the video has none in that language, get_transcript answers language_unavailable with the availableLanguages it does have, and charges nothing.
• source: "audio" transcribes what is said instead of reading captions — the way to get the original words when every caption is a translation. It is priced like any audio transcription (3 AI Units per minute of video, on top of the Cloud Request) and passing it is the consent.

list_languages answers the same question for free, from what is already stored.

One call across a channel

search_transcript accepts a list, which is how you answer "what has this channel said about X" without a round trip per video. Get the ids from list_channel_videos first:

{ "video": ["VIDEO_ID_1", "VIDEO_ID_2", "VIDEO_ID_3"], "query": "pricing" }

Each new video is billed at the usual 1 Cloud Request, and one unavailable video is reported in its own row rather than failing the call — the others were fetched and charged for, so you still get them.

It reads the picture, not only the captions

analyze_video looks at slides, charts, code samples and on-screen text that is never spoken aloud. ask_video then answers against that stored analysis, and every citation is checked before you see it: a visual claim has to match a frame that was actually recorded, a spoken claim has to land on a real transcript segment. Anything that fails is dropped, and when nothing survives the answer says the evidence is insufficient rather than producing a confident guess.

That is occasionally annoying — a refusal is a worse demo than a fluent answer — and it is the only version of this feature that is safe to put in front of an agent, because an agent repeats what it is told without the scepticism a human reader applies.

Auth, cost and limits

• If you pasted a token: Basic, not Bearer. The token is sent as-is; you do not base64-encode a user:pass pair. Clients that signed in carry their own credential and this does not apply.
• Verify your email first. Until you click the verification link every call returns 403 with {"error":"email_unverified"} — the most common first-call failure on a new account.
• Balances are shared with the REST API and the website. One Cloud Request is one new transcript; frame analysis spends 1 Cloud Request plus AI Units, and audio transcription spends AI Units on top of its Cloud Request. A run refused before it starts costs nothing, and analyze_video with estimateOnly: true quotes the price for free.
• Rate limit per minute, by plan: Free 60, Starter 200, Pro 500, Team 1,000 requests. The server is stateless, so a client re-runs initialize before every call and one tool call is several requests. analyze_video has its own ceiling of 10 starts per minute, shared with the REST route.
• RapidAPI tokens are refused here. That identity is metered per call and has no account behind it, neither of which survives a tool-calling session. Use a VidWords API token.
• Stateless by design. No resumable SSE streams, no session to delete; every tool answers in one shot. GET and DELETE return a JSON-RPC error rather than an HTML 404.
• No captions, or the wrong ones. For a video with no caption track, pass transcribeAudio: true to transcribe it from the audio; to replace captions that exist (all uploaded translations, say), pass source: "audio". Both are paid features, priced per minute of video, and never run without being asked for.

Full numbers: pricing.

Prompts

The server also publishes three prompts, which clients such as Claude show as ready-made actions. Each tells the model which tools to use and in what order, cheapest first.

Agent skill

skills/youtube-transcripts/SKILL.md is a drop-in agent skill for this server — tool selection, timecode spans, channel-wide search, the cost table and the error codes worth acting on, in the format Claude and compatible agents load directly.

Copy the folder into your agent's skills directory:

git clone --depth 1 https://github.com/haljishi/vidwords-mcp
cp -r vidwords-mcp/skills/youtube-transcripts ~/.claude/skills/

It assumes the MCP server is configured (see Quick start). The point of it is that an assistant which has read the skill knows to reach for search_transcript with a timecode span instead of pulling a whole two-hour transcript into its context.

Documentation

• Agent skill (SKILL.md)
• YouTube MCP server — overview
• Setup in Claude Code
• Setup in Claude Desktop
• Setup in Cursor
• Setup in ChatGPT
• Setup in Codex CLI
• REST API documentation

Support

Open an issue here for anything about the MCP surface — a tool that misbehaves, a client whose config we have not documented, a schema that could be clearer. Account and billing questions go to support.

License

The contents of this repository (documentation and configuration examples) are MIT licensed. The hosted service itself is proprietary and governed by the VidWords terms.

Independent product; not affiliated with YouTube or Google.