crawl_feed in Sume MCP: top-viewed posts are ranked within 3 pages

crawl_feed reads the recent or top-viewed items of a known public Instagram or TikTok account over Sume's hosted MCP. A call returns at most 24 items, from at most 3 pages. Because of that limit, a "top-viewed" answer is a ranking of the sample that was fetched, not a guarantee about the whole account. Treat it as "best of the latest pages" and say so in any report.

The server's own instructions say the same thing: play_count ranks the fetched sample, never guaranteed all-time.

For an agent report, the safe phrasing is "most viewed among the 24 most recent items that were read", with the sample size from sampled_count next to it. Anything stronger claims a fact that the tool did not establish.

Platform and kind

Each platform has its own kind values. Pick the one that matches the content you want.

Match the kind to the platform before you call. Asking for posts on TikTok or videos on Instagram is a wrong-shape request. The Instagram posts kind returns a mixed grid, while reels returns only reels.

Notice that the three cursor names differ. Instagram reels use max_id, Instagram posts use next_max_id, and TikTok videos use max_cursor. The result names the right one in cursor_parameter, so an agent that copies that name and the next_cursor value cannot mix them up.

Limits and what the result tells you

The limits are limit (default 10, maximum 24) and max_pages (default 3, maximum 3). The order can be recent or play_count. Items with a missing count have a null count and sort last. The answer reports sampled_count, pages_fetched, omitted_count and a next_cursor with a cursor_parameter, so an agent can tell how much was looked at and where to continue.

The numbers add up in a simple way: three pages is the ceiling, and 24 items is the ceiling for the list, so a request for more than 24 is not a deeper read. If you need more of the account, continue from the cursor in a second call and keep the two samples separate in your analysis.

Paging and the empty-feed trap

Cursors are strings. The cursor advances past all sampled rows, including the omitted rows, so pass back exactly what the result gave you under the named parameter. For TikTok, if a feed that you know is populated comes back empty, try the relevant two-letter region once. Instagram views exclude Facebook, and pinned reels may be absent from a list.

Omitted rows are items that were sampled but left out of the returned list. They still count toward the cursor, so the second call does not repeat them, and omitted_count tells you how many were skipped.

A request for the 10 most recent reels from a known account looks like this:

Add order: "play_count" when the question is about performance, and keep the default order when the question is about what the account posted lately. The two answers can be very different, and both are limited to the same sample.

If a call has to be split across several turns, store the cursor with the handle, the kind and the date. A cursor from one account or kind is not meaningful for another.

{
  "platform": "instagram",
  "kind": "reels",
  "handle": "example_handle",
  "limit": 10,
  "order": "recent"
}

From a list to a file

Each item has a permalink, media type, play count, publish time, and image and video candidate URLs. Those URLs may expire and are not hosted by Sume. To keep a video, pass its permalink to media-imports_create and wait for it; the lookup itself does not import or inspect anything. The group overview is on MCP tools and gates, and the connect steps are in the MCP quickstart.

The tool is bounded and unbilled discovery. It can finish within 30 seconds, and if it queues, the request_id is followed with jobs_wait and jobs_result, not by sending the lookup again.