OBSERVATORY

Registry observatory

How large each MCP server registry is, when it last returned new data, and how its catalog moved over the last 30 days — measured twice daily by the same sync jobs that build this site. Measurements, not endorsements: registries are listed in the order of their catalog size and nothing here ranks them.

mcptoplist.com is itself listed in some of these registries; we report measurements, not endorsements. Last measured Sep 4, 2026, 20:18 UTC.

COMPARISON

5 registries at a glance

Sum of catalogs 158,149 · 1 feed has returned no new data for more than 3 days
30-day trendStatus
Glama82,930+14,212
Sep 4, 2026
7 hours ago
Fresh
Official MCP Registry26,935+6,881
Sep 4, 2026
just now
Fresh
PulseMCP22,110−249
Aug 31, 2026
4 days ago
No new data
mcp.so20,919+7,699
Sep 4, 2026
7 hours ago
Fresh
Smithery5,255+565
Sep 4, 2026
7 hours ago
Fresh

Click a column to sort. The sum of catalogs double-counts servers listed in more than one registry; the deduplicated total is on the homepage. Definitions are in the methodology section below.

PER REGISTRY
GlamaFull crawl
Fresh
Listed servers
82,930
30-day change
+14,212
Aug 5, 2026Sep 4, 2026
Status
Synced 7 hours ago.
Last sync
Sep 4, 2026, 13:51 UTC
Newest published listing
Sep 2, 2026, 14:29 UTC
Newest first-time listing
Sep 4, 2026, 13:51 UTC
Known caveats (observed)
  • The public API has required an API key since 2026-08-26. No data was collected between 2026-08-26 and 2026-09-04 while a key was provisioned; the catalog series is flat over that period for that reason.
  • The catalog roughly doubled between late May and late June 2026 (23.7k to 47.6k). This tracked real crawl-coverage growth on Glama's side, mostly GitHub repositories not listed elsewhere.
Official MCP RegistryIncremental (changed listings)
Fresh
Listed servers
26,935
30-day change
+6,881
Aug 5, 2026Sep 4, 2026
Status
Synced just now.
Last sync
Sep 4, 2026, 20:15 UTC
Newest published listing
Sep 4, 2026, 13:03 UTC
Newest first-time listing
Sep 4, 2026, 13:08 UTC
Known caveats (observed)
  • Synced incrementally: each run fetches listings updated in the preceding two hours. The last-sync time reflects the newest changed listing the feed returned, not the run time, so it can lag on quiet days.
  • Delisting follows the registry's own status field. A listing that stops changing stays counted; only listings the registry marks deleted are removed.
PulseMCPFull crawl
No new data
Listed servers
22,110
30-day change
−249
Aug 5, 2026Sep 4, 2026
Status
No new data observed since Aug 31, 2026 (more than 3 days).
Last sync
Aug 31, 2026, 13:09 UTC
Newest published listing
May 21, 2026, 00:00 UTC
Newest first-time listing
Jul 17, 2026, 01:08 UTC
Known caveats (observed)
  • The public v0beta API returned at most 13,500 servers per crawl (observed 2026-05 through 2026-08), so the counts shown are a floor of the registry's catalog.
  • The v0beta API was retired on 2026-09-01 (every request now returns HTTP 410). Listings are frozen at the last successful crawl until the v0.1 API credentials are in place; the freshness figures above reflect that.
mcp.soFull crawl
Fresh
Listed servers
20,919
30-day change
+7,699
Aug 5, 2026Sep 4, 2026
Status
Synced 7 hours ago.
Last sync
Sep 4, 2026, 13:10 UTC
Newest published listing
Sep 4, 2026, 13:10 UTC
Newest first-time listing
Sep 4, 2026, 13:10 UTC
Known caveats (observed)
  • No public API. Counts come from the site's public sitemap and listing pages, with per-server details read from each listing page.
  • Between 2026-07-09 and 2026-08-25 a site redesign changed the page layout and the crawl observed no listings; the count stayed at exactly 13,170 for 47 days. That plateau is a measurement artifact on our side, not a change in the registry.
SmitheryFull crawl
Fresh
Listed servers
5,255
30-day change
+565
Aug 5, 2026Sep 4, 2026
Status
Synced 7 hours ago.
Last sync
Sep 4, 2026, 13:29 UTC
Newest published listing
Sep 3, 2026, 22:48 UTC
Newest first-time listing
Sep 4, 2026, 13:29 UTC
Known caveats (observed)
  • The unfiltered listing endpoint returns at most 500 results per crawl, so the catalog is enumerated per namespace (about 2,700 namespaces seeded from earlier crawls). Servers in a namespace not yet seen are picked up on a later run; counts are a floor.
  • Daytime crawls have been rate-limited (HTTP 429) since 2026-06; the crawl throttles adaptively and may land a partial result under its time budget, which shows as a small dip that recovers on the next run.
METHODOLOGY

What each number means

Listed servers
The number of live listings our crawl holds for the registry, after the registry's own duplicates are collapsed (two listings of the same repository count once). A listing is live under the same rule the leaderboard uses: registries we sync incrementally keep a listing until the registry marks it deleted; registries we re-crawl in full keep a listing for 30days after our last sighting of it. Where a registry's public API returns a bounded page of its catalog, the count is a floor — see that registry's caveats.
30-day change and trend
The series is reconstructed day by day from when each listing was first and last seen, using the liveness rule above, so its newest point equals the current count. A flat segment can be the registry or our crawl — the caveats say which when we know.
Last sync
The newest sighting our crawl wrote for the registry, on our clock. For incrementally synced registries this is the newest changed listing the feed returned, which can trail the run time on quiet days.
Newest published listing
The newest publish or creation date the registry itself reported for a listing, among everything our crawls fetched in the last 45 days — the registry's clock, not ours, capped at the last sync. It shows how recent the newest listing the registry returned is, which is a different question from when we last fetched it.
Status
Fresh: a sighting within 3 days. No new data: none for longer than that — the page reports the date and nothing more. Unchanged: crawls keep landing but no listing has been seen for the first time in 14days, which is what a crawl returning an identical catalog looks like. Freshness windows count back from the registry's own newest sighting, never the wall clock, so a paused feed freezes its numbers rather than draining them.

Caveats are hand-maintained, dated notes about behaviour we observed while crawling; they describe our measurement, not the registry. The registry list, cadences and deduplication method are on /sources; how the ranking itself works is on /methodology. Registry operators who spot a measurement error can write to hello@biff.ai.