Findings · 2026-10-15 · By VSNARY | Emmanuel Orta · 0 views
MCP reliability, October 2026: 6,000 remote MCP servers, 39,484 handshakes
The first monthly reliability edition for the remote servers in the official MCP Registry. Every server that has ever answered the handshake is contacted in rotation; no tool is ever called.
Between 8 and 15 October 2026, 98.7% of 39,484 handshakes with 6,000 remote MCP servers in the official registry were answered (95% interval 98.5–98.8%). The median endpoint answered in 1.17 seconds; the slowest 5% of answers took over 5.3 seconds. 53 endpoints went down at least once, and 689 of 3,498 endpoints that publish a tool list (19.7%) changed it.
Disclosure: CrawlCheck publishes this edition and sells checks built on the same measurements. No site, server or company is named in this post; the per-agent and per-server readings are on the live record linked below, regenerated and signed.
An agent that connects to a remote MCP server is trusting that the server answers, and that its tools are the ones the agent saw last time. This edition measures both for every remote endpoint in the official MCP Registry that has answered the handshake at least once. Each one is contacted in rotation: initialize, the initialized notification, tools/list, then the session is closed. No tool is ever called.
The figures #
| Measure | Value |
|---|---|
| Endpoints observed (publishers) | 6,000 (4,985) |
| Handshakes made | 39,484 |
| Answered (completed, or asked for credentials) | 98.7% (95% interval 98.5–98.8%) |
| Asked for credentials | 16,422 of 38,957 answers |
| Failed: unreachable / HTTP error / redirected / not MCP | 349 / 164 / 13 / 1 |
| Endpoints that went down (two failed observations in a row after answering) | 53 |
| Endpoints answering every observation | 5,201 of 5,590 with enough observations |
| Time to answer, median of endpoint medians / 95th percentile of answers | 1,170 ms / 5,275 ms |
| Endpoints whose tool list changed | 689 of 3,498 (19.7%), 1,128 changes |
A changed tool list is the figure to watch #
One endpoint in five that publishes its tools changed the list during the week. Most changes are ordinary releases. For an agent the point is narrower: a tool approved last week may not be the tool answering today, so the list is worth checking before a call, not once at install.
What this does not show #
- One vantage network: an endpoint unreachable only from Cloudflare's network reads as down.
- Endpoints that never answered are outside the panel.
- Observations are in rotation, so an outage shorter than the gap between two of them is not seen.
- October is a partial month: observations began on 8 October. The edition seals after the month closes.
The report: /registry/mcp/report · per-server readings: reliability index · monthly editions (Atom): feed · every edition: /editions.
Every figure above came out of CrawlCheck.
Point it at your own domain and see the same measurements, free.
The main product
Found this on your own site? We fix it for $749.
Scan free to see where you stand. The fix is one site, every finding implemented and re-measured, with a sealed before and after.
Questions this post answers
How reliable are remote MCP servers?
Between 8 and 15 October 2026, 98.7% of 39,484 handshakes with 6,000 remote servers in the official MCP Registry were answered; 53 endpoints went down at least once.
How fast do MCP servers answer?
The median endpoint answered the initialize request in 1.17 seconds; the slowest 5% of answers took over 5.3 seconds.
Do MCP servers change their tools?
689 of 3,498 endpoints that publish a tool list (19.7%) changed it during the week, 1,128 changes in all.
Does CrawlCheck call the tools?
No. It runs the handshake and lists tools, then closes the session.
Related findings
Comments
Comments are read before they appear. Nothing is published automatically, and no account is needed.
Writing about this? Facts, live figures and marks — every number on that page is dated and traceable to a scan.