YouTube MCP server: upload, read stats and reply to comments
A YouTube MCP server lets an agent upload a video, read a video's stats and reply to comments on your channel, through the official YouTube Data API. The quota, not the agent, is what limits it: 10 000 units a day, plus separate buckets of 100 uploads and 100 searches. Checked on November 16, 2026.
What a YouTube MCP server can do
YouTube is three surfaces for an agent, and only three. Publishing: upload a video file, set its title, description, tags, category, privacy and thumbnail. Engagement: read the comments on your videos, reply to them, hide them, delete them. Analytics: read view, like and comment counts per video, and the channel's subscriber count.
There is no fourth surface. YouTube has no direct messages, so an agent gets no inbox here, unlike the Telegram server or Instagram.
On PlugKit the tools are not YouTube-named, which surprises people reading the tool list for the first time. The server exposes 103 tools and three workflows as of today, and they are cross-platform: create_post and publish_post with a YouTube account id publish a video, upload_media carries the file, get_comments and reply_to_comment handle engagement, hide_comment and delete_comment moderate, and get_post_analytics and get_account_analytics read the numbers. One tool per action, a platform field, rather than eight near-identical tools per network.
That choice matters more than it sounds for an agent. A model asked to "reply to the new comments on my last video" picks get_comments and reply_to_comment on the first try. Faced with youtube_list_comment_threads next to instagram_get_comments, it starts guessing which one your account needs.
The quota is the real constraint
Every call against the YouTube Data API spends quota units, not requests. The default allocation is 10 000 units a day, and the costs are not flat. These are read today from Google's quota cost table.
| Call | What it does | Units |
|---|---|---|
videos.insert |
Upload a video | 1, from a separate bucket of 100 a day |
search.list |
Search videos | 1, from a separate bucket of 100 a day |
videos.list |
Read a video and its statistics | 1 |
channels.list |
Read the channel | 1 |
commentThreads.list |
Read comment threads | 1 |
comments.list |
Read replies | 1 |
playlistItems.list |
Read a playlist | 1 |
videos.update |
Change title, description, tags | 50 |
commentThreads.insert |
Post a top-level comment | 50 |
comments.insert |
Reply to a comment | 50 |
captions.insert |
Upload captions | 400 |
Reads cost one unit. Writes cost fifty. That ratio is the whole budget conversation: an agent that reads your comments all day will never touch the ceiling, and an agent that answers them will hit it at 200 replies. Edit the metadata on 200 videos and you have spent the same day's allowance without publishing anything.
Uploads and search have their own pools of 100 calls a day each, which is why a channel can publish daily forever and a discovery feature built on search.list dies in an afternoon. The counter resets at midnight Pacific time, not yours, so a scheduler firing at midnight in Paris is spending the previous day's budget. The YouTube API key guide covers how to request an extension when 10 000 units stops being enough.
Two defaults that will bite your agent
Privacy defaults to public. The videos.insert documentation states that if you do not set status.privacyStatus, the video is public, though some accounts default to private based on creator settings. An agent that uploads a draft without being told to set privacy has published it. Make the privacy value explicit in your prompt or your tool wrapper, and have the agent default to private when nobody said otherwise. PlugKit accepts public, unlisted or private on the post, so there is a field to set.
The comment scope is separate. Uploading needs youtube.upload, reading analytics and comments needs youtube.readonly, but replying to a comment or hiding one needs youtube.force-ssl. On PlugKit, channels connected before that scope was requested have to be reconnected before moderation works. If your agent uploads happily and fails on every reply, this is why, and no retry will fix it.
Titles are the third small trap: YouTube requires one and caps it at 100 characters. An agent that writes a long first line and lets it fall through as the title gets a rejection or a truncation, so pass the title explicitly.
What a YouTube MCP server cannot give you
Watch time, average view duration, impressions, click-through rate and traffic sources are not in the Data API. They live in the YouTube Analytics API, a separate product with targeted queries, alongside a Reporting API for bulk exports. Any MCP server advertising "YouTube analytics" that only talks to the Data API is giving you views, likes and comments, which is a fair amount but not what a creator means by analytics.
PlugKit is explicit about this line, and the limits are worth repeating because they decide whether the page you are reading is useful to you:
| What you want | Available through the MCP server | Where it really comes from |
|---|---|---|
| Views, likes, comments per video | Yes, totals since publication | Data API |
| Subscriber count | Yes, measured daily | Data API |
| Day-by-day history before you connected | No | Needs a separate Google audit |
| Watch time, retention, traffic sources | No | YouTube Analytics API |
| Direct messages | No | YouTube has none |
Videos you uploaded outside the server still show up in the analytics, which is the behaviour you want: connect a channel with 400 videos on it and the agent can read all of them. A channel that hides its subscriber count returns null rather than zero, so handle that in whatever you build on top. Social media analytics API goes platform by platform on why two tools never report the same number.
Connect it to Claude, ChatGPT or Cursor
PlugKit's hosted MCP server runs at https://api.plugkit.co/mcp over Streamable HTTP, stateless, nothing to install.
In Claude Code:
claude mcp add -t http plugkit https://api.plugkit.co/mcp
In Claude's apps it is Customize, then Connectors, then Add with the same URL. ChatGPT takes it as a custom MCP server, and Cursor and VS Code take a one-click link or a few lines of JSON. Authentication is OAuth 2.1 with PKCE and no client secret, with mcp:read covering profiles, accounts, posts and analytics, and mcp:write adding publishing, replying and account connection. API keys work too, as a bearer token, but they carry full access with no scopes.
Start on mcp:read for a week. Reading your comments is reversible. Publishing a video to a subscriber base is not, and neither is deleting a comment.
Write your own, or use a hosted one
If YouTube is the only channel you care about, write it. One Google Cloud project, one OAuth client, one refresh token, and the Data API is a well documented REST surface. A single-platform MCP server in front of it is a weekend, and you keep your own quota.
The arithmetic changes at the second platform. Every network ships a different OAuth flow, a different review, a different token lifetime and a different idea of what an analytics number means, and that work is per network, not per tool. MCP or REST API for your agent is the longer version of that trade-off, and the Instagram server shows what the same template looks like on a platform where Meta's app review is the gate.
A hosted server does not remove YouTube's quota. Nothing can: the quota belongs to whoever owns the API project. What it removes is the credential work, the token refresh and the per platform differences in how you publish. PlugKit is $49 a month for 5 connected accounts, cancel anytime, or $97 for 30. Seven days free, card required at sign-up, nothing charged before day 7.
FAQ
Is there an official YouTube MCP server from Google?
No. Google publishes the YouTube Data API, the Analytics API and the Reporting API, but no MCP server for them. Every YouTube MCP server is a third-party wrapper around those APIs, so the question to ask is which endpoints it covers and whose quota it spends.
Can Claude upload a video to YouTube?
Yes, through an MCP server that holds a YouTube OAuth token with the youtube.upload scope. Claude cannot reach YouTube on its own. With the server connected, create_post with a video media item and a title under 100 characters uploads it, and you should set the privacy value explicitly because the API default is public.
How much YouTube quota does an AI agent use?
Reads cost 1 unit and writes cost 50, against a default of 10 000 units a day. So an agent reading comments and stats uses almost nothing, while one replying to comments runs out at about 200 replies a day. Uploads come from a separate bucket of 100 calls a day.
Can a YouTube MCP server read watch time?
Not through the Data API, which only exposes views, likes and comments. Watch time, average view duration and traffic sources come from the YouTube Analytics API, a different product. Check which API a server talks to before you buy it for reporting.
Why does my agent fail when it replies to a YouTube comment?
Almost always the scope. Uploading uses youtube.upload and reading uses youtube.readonly, but replying, hiding and deleting comments need youtube.force-ssl. A channel connected before that scope was requested has to be reconnected, and retrying the call will not help.
One plug for every platform your agent touches
$49 a month for 5 connected accounts, cancel anytime. Every feature in every plan.
Get your API key →