// claude + superagnt_

Can Claude watch YouTube videos?

Not by itself: Claude's file uploads take documents and images, not video, and it has no built-in YouTube tool. Connect the superagnt_ MCP server and Claude reads any captioned video's transcript, comments and stats from its link, then summarizes, quotes or checks it.

Reading the captions is how Claude “watches”: it gets every line with its timestamp. The run below pulled one video's details, its 45 caption lines and its top comments in three calls, 9.5 seconds in all.

Copy the setup prompt.
Your agent does the rest.

works with any agent+ 8 more

Connect the superagnt MCP server at https://mcp.superagnt.com/mcp/youtube to this agent for me (I'll sign in through my browser). The setup guide for each kind of agent is at https://www.superagnt.com/install?facet=youtube

or follow the detailed setup instructions

» a new workspace starts with the YouTube tools already on. On an existing workspace, your agent switches them on during setup.

01// on its own, and with superagnt_

What Claude does alone, and what changes once it's connected

job	Claude on its own	with superagnt_
What was said	Nothing from a link. You can paste a transcript in yourself.	`data_youtube_videos` with `subtitles` returns every caption line with its start time, for any video that has captions.
Title, views, length	Only what web search finds about the video.	The details action: title, channel, views, duration, publish date, description and keywords.
What viewers think	No tool for comments.	The comments action: author, text, likes and reply counts, 20 a page with a token for the next.
Find videos	Web search.	`data_youtube_discover` searches YouTube by keyword, upload window, language and country, and reads trending.

02// a real run

The job, run on superagnt_

Real tool calls on superagnt_'s staging environment, the same calls Claude makes once the server is connected. Each result is what the client received, trimmed where it says so and otherwise unchanged, and each wall time was measured around the call.

the ask

“Watch IBM's “What is MCP?” video and tell me what it covers and how people reacted.”

01Read the video's details2.8 s wall timeOct 26, 2026, 12:39 UTC

→ calldata_youtube_videos

{
  "action": "details",
  "video_id": "eur8dUO9mvE"
}

← result

video_id: eur8dUO9mvE
title: What is MCP? Integrate AI Agents with Databases & APIs
channel_name: IBM Technology
channel_id: UCKWaEZ-_VweaEx1j62do_vQ
views: 775777
duration: 226
created_at: 2025-02-19T04:00:59-08:00
category: Education
type: NORMAL
keywords: IBM; IBM Cloud

» Trimmed: the video's description and the call's credit line.

02Read what was said3.9 s wall timeOct 26, 2026, 12:39 UTC

→ calldata_youtube_videos

{
  "action": "subtitles",
  "video_id": "eur8dUO9mvE"
}

← result

| start | text |
| --- | --- |
| 0.4 | If you&#39;re building AI agents, you&#39;ve probably heard about MCP or Model Context Protocol. |
| 5.24 | MCP is a new open source standard to connect your agents to data sources such as databases or APIs. |
| 11.68 | MCP consists of multiple components. |
| 13.94 | The most important ones are the host, the client, and the server. |
| 17.22 | So let&#39;s break it down. |
| 19.53 | At the very top you would have your MCP host. |

45 subtitles

» Trimmed: 39 of the 45 caption lines, and the call's credit line.

03Read the comments2.8 s wall timeOct 26, 2026, 12:39 UTC

→ calldata_youtube_videos

{
  "action": "comments",
  "video_id": "eur8dUO9mvE"
}

← result

| author | author_channel_id | text | likes | replies | created_at |
| --- | --- | --- | --- | --- | --- |
| @NeoReeo | UCe16PPteyepxeM-T-cTjzNQ | totally forgot to focus on the content and was figuring out how this guy is writing backwards | 61 | 3 | 1 year ago |
| @alexferrercampo7438 | UCE8QKAWI404GaIg39gtL-hQ | You made something simple into something complicated | 63 |  | 7 months ago |
| @devhims | UCDWd_XggrkR57wf2cQSA3tQ | In essence, MCP standardizes tool use for agents (augmented LLMs). | 132 |  | 1 year ago |

20 of 355 comments · next: continuation_token=Eg0SC2V1cjhkVU85bXZFGAYy8QIKxwJn…

» Trimmed: 17 of the 20 comments on the page, the end of the paging token, and the call's credit line.

Three calls, 9.5 seconds, and Claude has the video's numbers, every caption line of its 3 minutes 46 seconds, and the first page of its 355 comments, without anyone opening YouTube. Captions keep YouTube's HTML escapes (' for an apostrophe); Claude reads them as text.

03// set it up

Set it up in your client

Claude

01Open the prefilled Add custom connector dialog on claude.ai. One connector covers Claude on the web, Claude Desktop and Cowork.
02Click Add, then approve on the superagnt_ consent screen. It creates your superagnt_ account if you don't have one.
03Ask: “Watch this video and give me the five main points with timestamps: <YouTube link>”

Free plans can add one custom connector. On Team and Enterprise plans, an owner adds it under Organization settings, Connectors (Anthropic).

Claude Code

01In a terminal, run claude mcp add --scope user --transport http superagnt https://mcp.superagnt.com/mcp/youtube, then claude mcp login superagnt. The login opens a browser to approve; no API key needed.
02Start a new session.
03Ask: “Watch this video and give me the five main points with timestamps: <YouTube link>”

ChatGPT

01In ChatGPT on the web, go to Plugins, select the plus button, then Add custom MCP server.
02Name it superagnt, enter the server URL under Connection, and set Authentication to OAuth. The server URL is https://mcp.superagnt.com/mcp/youtube.
03Tick "I understand and want to continue", then select Create as a plugin. Sign in to superagnt_ when asked. No token needed.
04Install the plugin, then start a new chat, type @ and select superagnt.
05Ask: “Watch this video and give me the five main points with timestamps: <YouTube link>”

ChatGPT takes custom MCP servers on the web, not on mobile. Business and Enterprise/Edu workspaces get every tool; Pro accounts get read and fetch actions, which is all this job needs (OpenAI).

Codex

01In a terminal, run codex mcp add superagnt --url https://mcp.superagnt.com/mcp/youtube, then codex mcp login superagnt. The login opens a browser to approve.
02Start a new Codex session.
03Ask: “Watch this video and give me the five main points with timestamps: <YouTube link>”

04// what it can't do

What “watching” covers, and what it doesn't

It reads what was said, not what was shown. A slide or a screen recording with no narration comes through only as far as the captions describe it.
It needs captions: a video without any returns no lines to read.
Long videos come back long. A 30-minute talk had 692 caption lines: the compact view stops at 24,000 characters (461 lines) and says so, and the raw view returns all 692 at once. See How do I get a YouTube transcript into Claude?

05// what it reads

Everything Claude can read on YouTube through superagnt_

Tool	Actions	Good for
`data_youtube_videos`	details, subtitles, comments, recommendations	Summarizing, quoting and fact-checking one video, and reading its audience
`data_youtube_channels`	details, videos, shorts, search inside the channel, public contact email	Researching a creator before outreach, or tracking a competitor's uploads
`data_youtube_discover`	search, trending, audio details and the videos that use a sound	Finding what's been published on a topic this week