How many MCP servers are there?

Source · Marklode MCP registry census, October 2026 recount · as of 2026-10-17 VERIFIED

A registry counts listings. The same server published under several versions or entries is still one server, so a headline built on raw listings overstates the field several times over. This census counts the servers.

The September count, on 2026-09-12, found 31,309 servers behind 101,219 entries. Only those two figures were kept from it, so every breakdown below is from the October recount.

The Model Context Protocol (MCP) lets AI agents call tools and data through servers. The official public MCP registry lists those servers. Marklode reads every listing through the registry’s public API, then counts the distinct servers behind them.

The dedup rule

Registry entries are grouped by the server name the registry gives them, so the same server published under several versions or entries counts once. Each server’s figures come from its latest entry: the one the registry marks as latest, or, where none is marked, the last one listed. Its first-published month is the earliest publishedAt among all its entries.

A server has a public repository when its latest entry names one. How it is offered, its transports and its packages come from the same entry.

It is metadata only. We read what was published about each server, and nothing else. That limit is deliberate, and it shapes what the count can say.

Marklode’s API serves the live totals at api.marklode.com/v1/stats ↗. They move with each recount; this page keeps each recount as it stood.

Each table gives the servers in each category, their share of all 45,156 servers, and how many of them have no public repository.

T-01 Registry status

The status the registry gives each server’s latest entry.

Category	Servers	Share of all servers	No public repo
Active	44,613	98.8%	13,264
Deprecated	543	1.2%	119
All servers	45,156	100.0%	13,383

T-02 How a server is offered

From each server’s latest entry: a remote endpoint (an address an agent connects to), a package to install and run locally, both, or neither.

Category	Servers	Share of all servers	No public repo
Remote only	26,948	59.7%	12,391
Package only	15,557	34.5%	706
Remote and package	2,144	4.7%	212
Neither listed	507	1.1%	74
All servers	45,156	100.0%	13,383

T-03 Remote transport

Servers whose latest entry lists a remote endpoint of each transport type. A server listing both types counts under each, so the rows add up to more than the servers with a remote.

Category	Servers	Share of all servers	No public repo
Streamable HTTP streamable-http	28,356	62.8%	12,133
SSE (server-sent events) sse	1,085	2.4%	635
Rows added up	29,441	A server can count in more than one row

T-04 Package registry

Servers whose latest entry lists a package in each registry (npm, PyPI, OCI images and others). A server published to several counts under each, so the rows add up to more than the servers with a package.

Category	Servers	Share of all servers	No public repo
npm	11,344	25.1%	624
PyPI	4,506	10.0%	233
MCPB (MCP bundles) mcpb	1,210	2.7%	36
OCI (container images) oci	1,145	2.5%	45
NuGet	154	0.3%	7
Cargo (crates.io) cargo	80	0.2%	0
Rows added up	18,439	A server can count in more than one row

T-05 Servers first published, by month

Each server under the month of its earliest publishedAt across all its entries (UTC). That is when the registry recorded it, not when it was built. The last month is partial: the pull ended on 17 Oct 2026.

Month first published	Servers	Share of all servers	No public repo
Sep 2025	456	1.0%	69
Oct 2025	297	0.7%	79
Nov 2025	174	0.4%	52
Dec 2025	277	0.6%	59
Jan 2026	390	0.9%	44
Feb 2026	1,158	2.6%	127
Mar 2026	1,944	4.3%	188
Apr 2026	2,551	5.6%	377
May 2026	3,038	6.7%	399
Jun 2026	3,860	8.5%	706
Jul 2026	5,089	11.3%	1,452
Aug 2026	6,547	14.5%	2,231
Sep 2026	11,874	26.3%	3,715
Oct 2026	7,501	16.6%	3,885
All servers	45,156	100.0%	13,383

• It does not say any server is safe, or unsafe. We never ran one, so we can’t tell you what any of them does.
• It does not say how many servers are maintained, used, or used by anyone at all. A listing is a claim that something exists, not evidence that anyone depends on it.
• It does not rank servers or vendors.
• It does not say a server without a public repository is hiding anything. Plenty of honest software is closed source. It means you can’t read the code.
• It does not name servers. The files hold counts only: no server name, address or description. To look up one server, use Marklode’s checker.
• It covers one registry, the official public MCP registry. A server listed only somewhere else is not in it.
• And it is a snapshot. Registries change daily, which is why the census is recounted monthly and every figure carries the date it was checked.

• 2026-10.jsonJSON · 7,427 bytesEvery figure on this page in one file: the headline, the September count, each breakdown, the dedup rule and the limits.SHA-256 1dad04eeb345486d2ce9e050bfeac1ffd2392128eef3f8f52336362c0946910b
• 2026-10.csvCSV · 1,592 bytesThe same figures, one row each: table, category, unit, count, share of all servers, and how many have no public repository.SHA-256 dda944829dac3c4b8ff196c6acfed489655dc8c56df1194a329bdecd75e5a92d
• index.jsonJSON · 1,013 bytesEvery recount this site publishes, with each file’s size and SHA-256. Its own hash changes with each recount.SHA-256 6843a4e35d57c2f4e3d06ee7beccab9f1cf4030a1925e73b133c8a68a4fd661a

Check a download: shasum -a 256 2026-10.csv on a Mac, sha256sum 2026-10.csv on Linux. The result should match the line above.

Licence

The counts are published under Creative Commons Attribution 4.0 International (CC BY 4.0) ↗: use them, share them and build on them, with credit to “Marklode MCP registry census, Deimos AI LLC” and a link to this page.

Deimos AI LLC (Marklode). The MCP registry census, October 2026 recount, counted 2026-10-17. https://www.deimoscore.com/marklode/census/. Licence: CC BY 4.0.