Measurement: 8.6% of listed entries with a remote endpoint answer HTTP but not MCP

Hello — this is a measurement report, not a bug report or a feature request.
I have been probing the registry nightly and thought the numbers might be
useful. No action requested.

Method

Every night at 03:00 UTC I fetch the full registry, then send a plain
JSON-RPC initialize to every entry that declares a remotes[] endpoint,
followed by tools/list. Failures are recorded with a reason, so "refused
us" is never conflated with "is not there". Both streamable-http and legacy
SSE are attempted, and three protocol versions are tried before an entry is
written off.

Results, 2026-09-07

28,139 registry entries, of which 16,305 declare a network endpoint. Every
one of those 16,305 was probed:

149,725 tools were recorded across the servers that answered.

The row worth your attention is the third. Those 1,401 entries respond over
HTTP but do not speak MCP at all: they return HTML, unrelated API errors, or
nothing parseable. They are not auth failures and not outages — both are
counted separately above.

Four of them are expired domains now serving for-sale parking pages, all
through the same ad network, while still listed as active:

• backlog-forge.arcanum-plus.com
• toolpipe.dev
• mcp.TEST-blockscholes.com
• mcp.tradallo.com

An agent configured from the registry connects to infrastructure the
original publisher no longer controls.

Two smaller observations

• Declared endpoints grew from 14,426 to 16,305 over the eight days measured
  — roughly 235 per day — so any absolute figure here ages quickly.
• 24.8% of servers that do speak MCP have no mcp substring anywhere in
  their URL path, across 3,026 distinct paths. Discovery heuristics keyed on
  the path will miss a quarter of working servers.

Reproducing this

Crawler, raw archived responses and analysis:
https://github.com/yassinht/mcp-observatory

Observations go into a signed RFC 6962 Merkle log, so these figures can be
recomputed rather than taken on trust. Corrections welcome — I have had to
correct my own numbers several times and the README records those.

Happy to share the full per-entry breakdown in whatever format is useful, or
to run a specific query against the archive.

Start by reading the measurement report and the linked mcp-observatory crawler, archived responses, and analysis. No repository file, test, requested change, or completion condition is identified; the report explicitly says that no action is requested.