A census of the agent-reachable web
The ecosystem built the servers and forgot the front door.
of the organisations that provably run an MCP server publish nothing an agent could use to find it. 4339 of 7421 measured.
How many big websites run an MCP server is a settled question: almost none, and Cloudflare measured it at 200,000 domains. This asks the other one. Of the organisations that demonstrably run a server, how many can an agent actually reach?
Why this is hard
The organisations above are not being careless. There is no ratified way to do this. SEP-2127, the proposal for server discovery over HTTP, has been open since 21 January 2026, is still labelled a draft, and missed its own target date of 3 April 2026. In July it was moved off the standards track altogether and refactored onto the extensions track, where Server Cards are being incubated rather than standardised.
The design also moved while people were adopting it. The card left the domain root for a path relative to the MCP endpoint, and even that is only a recommended location: cards may live at any unreserved URI. Discovery at the level of a domain was handed to a separate cross-ecosystem effort, the Linux Foundation AI Catalog, which needs a vote from both the A2A and MCP steering committees before it means anything. The experimental extension repository still says in as many words that it is not an accepted or official MCP extension.
The extension repository tracks what SEP-2127 is waiting on. Twelve items are labelled as blockers; ten are closed and two are still open: best practices for client implementors, and reference implementations in Python and Go. Both leads of the working group that owns the proposal reach the end of their terms on 14 August 2026.
Checked against the pull request, its labels, the working group charter and the extension repository's issue tracker on 7 August 2026. Every claim here is one click from its source.
The most-used path is in no specification
The path most organisations use is /.well-known/mcp/server-card.json, on 971 domains, and it appears in no specification document. It spread through blog posts instead. The AI Catalog, where domain-level discovery is actually heading, answered at its
well-known path on 258. That is a floor, not its adoption: its own
specification makes that path optional and lets a catalog live at any URL.
How to read this
Domains we were not permitted or not able to measure are excluded from every number above. A domain that blocked our crawler is reported as its own category and never counted as a zero.
The population is organisations the official MCP Registry proves run a server. They have already done the harder part, so if reachability is low here it is lower everywhere else.
Read the methodology · Get the data · See every domain
Last complete run 13 September 2026. Methodology 0.5.0.
The ecosystem keeps growing. Reachability has not followed.
Servers in the official MCP Registry, cumulative. This is the registry's own count, not ours, and a different question from whether an agent can find any of them.
787 servers in 2025-09, 65,787 by 2026-08. Monthly additions went from 787 to 18,952 in the last full month. Against that, 23% of the organisations the registry proves run a server answer a handshake from their own domain.
Show these numbers as a table
| Month | New servers | Cumulative |
|---|---|---|
| 25-09 | 787 | 787 |
| 25-10 | 713 | 1,500 |
| 25-11 | 750 | 2,250 |
| 25-12 | 829 | 3,079 |
| 26-01 | 1,238 | 4,317 |
| 26-02 | 2,952 | 7,269 |
| 26-03 | 6,180 | 13,449 |
| 26-04 | 7,961 | 21,410 |
| 26-05 | 9,916 | 31,326 |
| 26-06 | 12,774 | 44,100 |
| 26-07 | 18,952 | 63,052 |
| 26-08 partial | 2,735 | 65,787 |