Every "how many MCP servers exist" number I could find counts registry entries. The official Model Context Protocol registry keeps every historical version of every server, so the entry count and the server count are different things. I pulled all of it and did the arithmetic.
Method
- Source: the public registry API, every record by cursor pagination, on 14 September 2026.
- Deduplication: one record per server name, keeping the latest version.
- Hosts: parsed from the declared remote URLs.
- Metadata only. I never connected to a server, so "declares an HTTPS endpoint" is not the same as "up" or "safe".
The numbers
| Figure | Value |
|---|---|
| Raw registry entries | 101,219 |
| Distinct servers (latest version per name) | 31,309 |
| Entries that are superseded versions | 69% (3.23x inflation) |
| Servers with no source repository | 7,090 (22.6%) |
| Remote-only servers with no repository | 6,484 (20.7%) |
| Remote / package / both / neither | 18,984 / 13,427 / 1,516 / 414 |
| Servers sharing a hostname with 50+ others | 5,460 (17.4%) |
| Top three hosts combined | 4,525 servers (14.5% of everything) |
| Distinct publishers | 18,500 |
| Publishers with exactly one server | 16,594 (89.7%) |
| Marked deprecated | 334 |
| New servers first seen, Sep 2025 | 456 |
| New servers first seen, Aug 2026 | 6,673 |
What this means if you are choosing servers for an agent
Version inflation. Any count of registry entries overstates the ecosystem by 3.2x. If a directory, a benchmark or a pitch deck says "100k MCP servers", it counted versions.
Auditability. One server in five is remote-only with no repository. There is no code to read; the only way to find out what it does is to connect to it. If your policy is "review before you allow", a fifth of the registry is unreviewable by construction.
Concentration. 17.4% of servers share a hostname with at least fifty others, and three hosts front 14.5% of the entire registry. Some of that is legitimate gateways. Some looks like bulk registration. Either way, "which host is this really talking to" is a better question than "which publisher is listed".
Growth. New servers per month went from 456 to 6,673 in a year. Whatever review process you have today will not scale to next year's registry by hand.
Check a server before you trust it
- Live totals, no key needed: https://api.marklode.com/v1/stats
- Look up one server name or one publisher namespace (free tier, 1,000 calls a month): https://rapidapi.com/thearchitect0x/api/marklode-verify
- Pull the whole deduplicated dataset yourself: https://apify.com/thearchitect0x/mcp-registry-scraper
I refresh the census monthly and will post the deltas. Corrections and questions about method are welcome in the comments.
Top comments (1)
The 69%-superseded inflation explains every bloated "MCP ecosystem" headline I have seen. The 22.6% with no source repo is the number that should worry people more than it does: an agent pointed at a remote-only server with no repository is trusting a hostname with a tool list. Also appreciated the "metadata only, never connected" method note - counting without touching is the right way to run this census.