If you use ChatGPT, Claude, or Grok to think about YouTube videos, you know the loop: switch tabs, paste a video URL into a transcript tool, wait, copy the output, switch back, paste it into your chat, then finally ask your real question. Six clicks of context switching before you do the work you actually opened the chat to do. The Model Context Protocol (MCP) makes that loop disappear.
Anthropic published MCP in late 2024 as an open standard for connecting AI chat clients to external tools. Through 2025 and 2026, OpenAI and xAI shipped their own MCP integrations. Claude, ChatGPT, and Grok now each expose a path for remote MCP connectors, with different plan and setup requirements per vendor. We run one at https://api.youtubevideotranscript.io/mcp; once it is connected, your chat fetches YouTube transcripts on its own, answers using the content, and charges your account’s credits the same way the web app does. See our MCP page for a quick overview and links into the per-client install guides.
What MCP actually is
Think of MCP as the equivalent of HTTP, but for chat-client tools. It is a small JSON-RPC-style protocol that describes how a chat client can discover a server’s tools, call them with typed arguments, and receive structured responses. The tools themselves are server-defined. Our server exposes three:
- get_transcript: fetch one video’s transcript by URL. Charges 1 transcript credit on success. Supports a
formatparameter (text, timestamps, or json) so the model can pick whichever shape costs the fewest tokens for the user’s actual question. - list_channel_videos: enumerate every video on a channel, page by page. No credit cost.
- list_playlist_videos: same for playlists. No credit cost.
That is intentionally a small surface. Bulk channel jobs (where we fetch hundreds of transcripts as one batch) stay on the web app and the public REST API: chat clients cannot reasonably keep a connection open for the minutes that batch takes. MCP is purpose-built for interactive use.
Why this beats copy-pasting transcripts into your prompt
Copy-pasting works. It is also error-prone, slow, and changes the kind of question you ask. You only paste a transcript once, so you front-load a single big prompt and hope you asked the right thing. You stop asking follow-ups when the transcript is too long to re-paste. You lose the timestamps because the formatting got stripped during the copy. The output you do get is harder to cite back to the original because the chat has no idea where the transcript came from.
MCP changes the shape of the interaction. The model itself decides when to fetch a transcript, what format it needs (plain prose for a summary; timestamped for “when did she say X?”), and how to cite results back to the video. You can ask three follow-up questions about three different sections of the same video and the model fetches once, caches in context, and answers them naturally. You can drop a channel URL and ask the model to look for a topic across the back catalog. None of this requires you to leave the chat.
There is a deeper differentiator worth being honest about: most “use YouTube transcripts in your AI chat” guides on the internet today are still describing the copy-paste workflow, or local stdio MCP servers you have to run on your own machine. Our endpoint is a remote MCP server with OAuth 2.1 and API-key auth, hosted by us, callable from any MCP client. No Node process to babysit, no localhost ports, no permissions to manage. Install once, works from any device signed in to your chat account.
Install in ChatGPT
ChatGPT exposes MCP through Apps & Connectors with Developer Mode enabled. ChatGPT requires Developer Mode on web. OpenAI currently lists Developer Mode for Pro, Plus, Business, Enterprise, and Education accounts. Free is not supported, and mobile apps are not documented for custom MCP app setup.
High-level steps: open Settings → Apps & Connectors, scroll to Advanced settings and toggle Developer mode on. Click Create, paste https://api.youtubevideotranscript.io/mcp as the server URL, choose OAuth, and click Create. Sign in on the consent screen that opens. In any new chat, toggle YouTube Transcript on via the + menu under Developer mode. See the full walkthrough on /docs/mcp/chatgpt including the workspace-admin step for Team and Enterprise installs and the per-conversation enable behavior that catches most people on first use.
Install in Claude
Claude splits into two install paths. The hosted surfaces (Claude.ai web, Claude Desktop, and the Claude mobile apps) connect via the Custom Connectors UI under Settings → Connectors. Anthropic only exposes OAuth here, so you click Add custom connector, paste https://api.youtubevideotranscript.io/mcp, click Add, and walk through our consent screen. Free, Pro, Max, Team, and Enterprise all support custom connectors; Free is limited to one.
Claude Code and the VS Code / JetBrains / Cursor extensions take a Bearer header directly. Run claude mcp add --transport http --scope user --header "Authorization: Bearer yvt_live_YOUR_KEY" youtube-transcripts https://api.youtubevideotranscript.io/mcp in your terminal and the server appears in /mcp inside any Claude Code session. The full Claude install matrix, including the .mcp.json file you can commit to a repo so your team picks up the same connector, is on /docs/mcp/claude.
Install in Grok
Grok exposes a custom connector flow at grok.com/connectors with a Bearer-token field, so install does not require OAuth or a Developer Mode toggle. If your account shows the Connectors menu, open grok.com/connectors, click New Connector → Custom, paste the MCP server URL, paste your yvt_live_* API key in the Bearer / Authorization field, and click Add. Availability can vary by Grok surface and account. The Grok integrated inside x.com / the X social app does not yet expose this menu; install at grok.com with the same X account instead. Full instructions plus the xAI Python SDK form for developers are on /docs/mcp/grok.
What to actually do with it once installed
Four prompts that show off what changes when your chat can pull transcripts on its own:
- Summarize a video without watching it. “Summarize this video in five bullets: [URL]”. The model picks the cheap
textformat under the hood, fetches once, and answers. Useful for triaging long podcast episodes or conference talks before deciding whether to watch. - Cite specific moments. “In [URL], when does she define recursion?”. The model picks the
timestampsformat, finds the line, and replies with the [mm:ss] anchor you can paste into a clipboard or a message. - Find clip-worthy moments for short-form content. “Read this 45-minute interview [URL] and give me three 30-second moments that would work as TikTok or LinkedIn clips.”. Useful for creators repurposing long-form content into shorts.
- Build a chapter list. “Generate YouTube-ready chapter markers for [URL] with one chapter per topic shift.”. Faster than scrubbing the timeline by hand, and the model can re-do it with different granularity if the first pass is too coarse.
For everything beyond ad-hoc exploration (full channel pulls, AI training datasets, structured CSV exports), see the companion post YouTube transcript for AI training. MCP and bulk export complement each other: MCP for one-video, in-chat thinking; bulk export for full datasets and offline processing.
Costs and plan requirements
Every transcript fetched through MCP charges your YouTube Video Transcript account exactly 1 credit, the same as a transcript fetched through the web app. The free tier’s 10 one-time credits work inside MCP, so you can install the connector, try it on three or four videos, and see if it is worth upgrading before committing to a paid plan. Channel and playlist listing tools are free.
The chat app’s own plan requirements are separate from ours and roll out at the chat vendor’s pace. ChatGPT requires Developer Mode on web; OpenAI currently lists Developer Mode for Pro, Plus, Business, Enterprise, and Education accounts, with Free unsupported and no documented mobile path. Claude works on Free with a one-connector limit and scales up on Pro / Max / Team / Enterprise. Grok availability depends on whether your account exposes the Connectors menu. For an overview of all YouTube Video Transcript plans see /pricing; for the full MCP documentation home with all three install guides linked from one place see /docs/mcp.
When MCP is not the right tool
MCP is built for interactive use. If you need 500 transcripts for a fine-tuning dataset, our bulk job system (web app or POST /api/v1/jobs) is the right tool: it processes the channel in the background and gives you a ZIP, instead of forcing your chat client to wait minutes on an open connection. If you need scheduled, unattended access from a server, an API key on the REST API is also better than MCP: no model in the loop, fully programmatic.
Concrete cases where the REST API beats MCP: pulling every upload from a multi-thousand-video research channel for an eval set, running a nightly cron that ingests new uploads from 50 podcasts into a vector store, feeding transcripts into a Zapier or n8n workflow, populating a Snowflake or BigQuery table for analytics, batch-translating captions for a content team. Anything where you would have been writing a script anyway is cheaper to do as a real script against /api/v1 than to drive through a chat conversation. See /docs for the full REST reference.
MCP shines when there is a human at a keyboard who would otherwise be tab-switching. That is the friction the protocol is designed to remove. Use it for the cases where you already have your AI chat open and you would have been pasting a transcript five minutes from now anyway.