Introducing SocialKit MCP — Give Claude, Cursor, and Any AI Agent Access to Every YouTube, TikTok, and Instagram URL

June 1, 2026•Jonathan Geiger

mcpai agentsclaudecursorannouncementdeveloper-tools

Published June 2026.

I just asked Claude "summarize this Instagram Reel and tell me whether the top comments are positive or negative" and pasted a URL. Claude pulled the transcript, fetched the comments, ran sentiment, and gave me back a one-paragraph answer. No coding, no copy-paste, no API wrappers — just the URL and a question.

That's what we shipped today: SocialKit now has a hosted MCP server at mcp.socialkit.dev.

What it is, in one line

The Model Context Protocol is Anthropic's standard for letting AI agents call external tools. We took every SocialKit endpoint — 33 of them across YouTube, TikTok, Instagram, Facebook, and direct video URLs — and exposed each one as an MCP tool. Connect once, and your AI gets:

youtube_transcript, youtube_summarize, youtube_stats, youtube_comments, youtube_channel_stats, youtube_videos, youtube_search, youtube_download
tiktok_transcript, tiktok_summarize, tiktok_stats, tiktok_comments, tiktok_channel_stats, tiktok_channel_video_metrics, tiktok_search, tiktok_hashtag_search, tiktok_download
instagram_transcript, instagram_summarize, instagram_stats, instagram_comments, instagram_channel_stats, instagram_channel_posts, instagram_channel_reels, instagram_reels_search, instagram_download
facebook_transcript, facebook_summarize, facebook_stats, facebook_comments, facebook_channel_stats
video_transcript, video_summarize (for direct video file URLs)

From the AI's perspective, those are just 33 new functions it can call. From yours, it means you can describe a social-media task in plain English and the agent figures out which tools to call and in what order.

Setup in 30 seconds

Claude Desktop

Open Settings → Developer → Edit Config and add:

{
  "mcpServers": {
    "socialkit": {
      "url": "https://mcp.socialkit.dev/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_ACCESS_KEY"
      }
    }
  }
}

Restart Claude Desktop. The 33 tools appear in the tools panel. Done.

Claude Code

claude mcp add --transport http socialkit https://mcp.socialkit.dev/mcp \
  --header "Authorization: Bearer YOUR_ACCESS_KEY"

Cursor

Open Settings → MCP → + Add new MCP server and paste:

{
  "socialkit": {
    "url": "https://mcp.socialkit.dev/mcp",
    "headers": {
      "Authorization": "Bearer YOUR_ACCESS_KEY"
    }
  }
}

You can grab a free access key at the SocialKit dashboard. Twenty free credits, no credit card.

The MCP server runs over HTTP-streamable transport, which means there's nothing to install locally — no npx -y keep-it-updated dance, no Python venv, no binary. It's a URL.

Three things you can do today

1. "Summarize this video and pull the engagement"

Summarize https://www.youtube.com/watch?v=dQw4w9WgXcQ and tell me how many views and likes it has.

Claude calls youtube_transcript, runs the summary through its own reasoning, then calls youtube_stats and tacks the engagement on. One prompt, two tool calls, one clean answer.

2. "Are the comments positive?"

Pull the top 50 comments from this Reel: https://www.instagram.com/reel/DHeBXm5tBLu/
and tell me whether the sentiment is positive, negative, or mixed.

instagram_comments returns the raw comments, the model runs sentiment over them itself, and writes you a paragraph. The model is the sentiment classifier — no separate /sentiment endpoint needed.

3. "Find me the top 10 cat videos this week"

Search TikTok for "cat compilation 2026", grab the top 10, sort by views, and tell me which channel each one is from.

tiktok_search returns the candidates; the model handles the sort + aggregation. The agent decided when to call which tool.

Why this works

Traditionally, building an "AI that understands social media" meant either (a) scraping with Puppeteer and explaining the resulting HTML to your LLM, or (b) writing function-calling glue code for every endpoint by hand. Both are brittle. Both don't compose.

MCP changes the math. Your agent gets the tools as first-class capabilities — discoverable, typed, callable in any order, composable with other MCP servers your agent has. The agent figures out the rest.

A few things specifically true about ours:

Hosted, not local. No npm install -g, no binary, no npx. Just a URL.
Same access key as the REST API. No new auth system — your existing access key works, costs the same per-call, hits the same rate limits.
Stateless. The server doesn't store your URLs, your responses, your key, or anything else. It forwards your request to api.socialkit.dev, returns the response, and forgets you existed.
Open source. github.com/geiger01/socialkit-mcp — self-host if you want to.

What we're using it for internally

A few internal flows we've already migrated:

Daily competitor digest. A scheduled Claude Code job hits youtube_videos on five competitor channels, summarizes any new uploads, runs comment sentiment on the top three, and posts the digest into Slack.
"What's this URL?" Bookmarklet that feeds any social URL to Claude with a fixed prompt — "tell me what this is in 3 lines." Uses whatever *_stats + *_transcript tool the URL belongs to.
Customer support triage. Agents reading a support ticket that includes a YouTube/TikTok URL automatically pull the transcript so we know what the user is referring to without watching the video.

None of these would have happened if each agent needed a hand-written HTTP wrapper.

What's next

The MCP exposes the current REST surface 1:1. Coming soon: bulk-endpoint tools (so agents can batch instead of looping), webhook-based monitoring tools (subscribe an agent to a channel and have it alerted on new content), and resource templates so agents can read SocialKit responses as MCP "resources" instead of just tool results.

If you build something cool with it, let us know. And if you hit a gap — a SocialKit endpoint you wish were exposed, or a workflow you wish were possible — open an issue on the GitHub repo. Adding a new tool is a 5-line change.

Getting started

Grab a free access key (20 credits, no card).
Paste the config above into your client.
Ask your AI to do something. Anything.

The MCP guide in our docs has the full surface — every tool, every parameter, troubleshooting tips. And the /mcp landing page has the rest.

Welcome to the era where your AI knows what's on the internet.