Skip to content
ProtocolProbe
Explore

Directory · changes · newest registry pull 2026-10-24

Changes

Every change we record, in one place: what changed in the official MCP Registry list between our pulls, what changed when we read the facts of servers again, and the servers that are no longer listed. Each pull has its own page with every name, and each re-check its own anchor on this page. The same records are in the API as JSON, CSV or NDJSON.

Registry changes

Each row is the change between two pulls of the official MCP Registry list: the day named and the pull before it (a monthly refresh compares two pulls further apart and is marked so). A server is new when it appears in the list for the first time, has a new version when its latest listed version differs from the pull before, is no longer listed when it was in the pull before and is not in this one, and is listed again when it had been in an earlier list, was not in the pull before, and is in this one. The registry gives no reason for a change, and we do not guess one.

On 2026-10-24 the list had 49,445 servers. We had read the tool list of 22,725 of them (355,690 tools), without calling any tool.

Every pull, oldest on the left

Servers in the registry list, per pull
49,50048,5002026-10-232026-10-242026-10-23: 48,551 servers2026-10-24: 49,445 servers

The axis starts at 48,500, not at zero, so small daily changes are visible.

New servers
10,00002026-10-232026-10-242026-10-23: 9,8032026-10-24: 928
New versions
10,00002026-10-232026-10-242026-10-23: 5,6462026-10-24: 1,906
No longer listed
25002026-10-232026-10-242026-10-23: 2292026-10-24: 34
Listed again
102026-10-232026-10-242026-10-23: 02026-10-24: 0

Each bar chart starts at zero and has its own scale (the top value is marked). A monthly refresh covers more days than a daily pull, so its bar is taller.

Only 2 pulls are stored so far (2026-10-23 to 2026-10-24). A trend over so few pulls says little; the charts will fill in as more days are loaded.

By month, newest first

MonthPullsNewNew versionsNo longer listedListed againServers at the last pull
October 20262 (1 monthly)10,7317,552263049,445 (2026-10-24)

Sums of the pulls of each month (UTC day of the pull). A monthly refresh compares two snapshots further apart, and its whole comparison counts in the month of its snapshot. A server that changed on several days counts once per day; a server can be new in one pull and no longer listed in a later one.

Per pull, newest first

Each pull has its own page with every name; click its date.

October 2026 · 2 pulls
PullCompared withServers listedNewNew versionsNo longer listedListed again
2026-10-2349,4459281,906340
(monthly)2026-10-0448,5519,8035,6462290

The same records as a file: CSV · NDJSON · JSON (add from and to for a range of up to 400 days; API docs).

Fact re-checks

When we read the public facts of servers again (package downloads, GitHub stars and repository state, the listed version against the package registry, dependency audits and known vulnerabilities), this section counts what changed for the same server and the same listed version. A server that listed a new version is not compared here: it is in the registry changes. Readings where our own lookup failed (a rate limit, a network error, a tool that could not run) are never counted as changes; they are left out and their number is given. A change is a difference between two readings of public records, not a judgement of the server or its publisher, and no reason is given or guessed.

Facts re-checked

free facts of the 33,102 servers unchanged since 2026-10-23, re-read 2026-10-24; all npm/PyPI dependency audits re-run 2026-10-24.

Compared with the facts published after the 2026-10-23 refresh (read 2026-10-06 to 2026-10-23); the facts compared now were read 2026-10-22 to 2026-10-24. Read again in this run: package, GitHub and OSV facts of 33,102 servers, version pins of 11,830 servers, npm dependency audits of 11,756 servers, PyPI dependency audits of 4,535 servers, repository languages of 33,102 servers. 48,551 servers had the same listed version in both and were compared (a fact that was not read again is the same reading on both sides). 410 readings where our own lookup failed are left out of every count.

  • Facts of the same server and the same listed version, read in two runs. A change is a difference between two readings of public records, not a judgement of the server or its publisher; no reason is given or guessed.
  • Our own failures (rate limits, network errors, a tool that could not run) are never counted as changes: 410 field readings were left out for that reason (ourLimit).
  • 0 servers listed a different version in the two runs and are not compared here (new version: see the registry changes).
  • Dependency audits: npm audit counts vulnerable packages in the resolved tree; the PyPI audit counts OSV advisories in the resolved tree. A count can move because the advisory database changed, without any change to the server.
  • Same listed versions on both sides: the rise in npm servers with a high or critical vulnerable dependency comes from the advisory database, which changed between the two reads (GitHub reviewed GHSA-mrhx-c7xc-m3w5, zod below 4.6.6, severity high, on 2026-10-23 18:26 UTC; read from api.github.com 2026-10-24), and from dependency versions resolved on the day of each audit. It is not a new release of these servers. The audit rows keep counts only, not advisory ids, so the share due to that one advisory is not measured.

Sources: our pulls of the official MCP Registry list (registry.modelcontextprotocol.io), compared by the daily load or a monthly refresh (the day is the UTC day of the pull, or the snapshot name of the refresh); and our readings of npm, PyPI, GitHub and OSV, compared between two refreshes. A change is a difference between two lists or two readings, not a judgement of the server or its publisher. We called no server's tools for this page.