CrawlCheck

Findings · 2026-11-14 · By

One in four MCP hosts changed its tools in four days. Nothing told the agent.

We re-read the tool list of every public server in the MCP Registry. Most changes rewrote a description, and most kept the same version number.

Between 10 and 14 November 2026, 706 of 3,167 hostnames in the MCP Registry (22.3%) served a different tool list than the one CrawlCheck read before. 83.9% of recent changes rewrote a tool description, and 71% of the latest 500 kept the same server version number. An agent approved against the earlier list is now using tools nobody re-reviewed.

The last 440 changes, by kind (a change can fall in more than one row)
A tool description was rewritten369 (83.9%)
A tool was added207 (47%)
A tool was removed36 (8.2%)
Only inputs or annotations changed36 (8.2%)

When you connect an AI agent to an MCP server, you approve a list of tools: their names, their descriptions and what they are allowed to do. That approval is a snapshot. The server can change the list the next day, and nothing in the protocol tells the agent or you.

We wanted to know how often that happens, so we counted. Every figure here was read on 14 November 2026.

What we counted #

We re-read every public server in the MCP Registry that answers without a login. We ask for the tool list and compare it with the previous read. We never call a tool.

  • 6,744 servers returned a tool list. 4,953 of those lists (73.4%) were not the list we read before.
  • That figure counts registry entries, and one publisher can list thousands of near-identical entries on one host. Counted by host, it is 706 of 3,167 (22.3%). By registered domain, 677 of 2,843 (23.8%).
  • The median gap between the two reads was 8.7 days.

So the fair headline is the smaller one: about one host in four changed its tools in four days.

What changed #

The last 440 changes, recorded on 13 and 14 November. A change can fall in more than one row.

ChangeServersShare
A tool description was rewritten36983.9%
A tool was added20747%
A tool was removed368.2%
Only inputs or annotations changed368.2%

The first row matters most. A tool's description is the text the model reads to decide when to call the tool and how. Rewrite it and the agent can behave differently, while everything the user sees stays the same.

The version number did not move #

We looked at the latest 500 changes, recorded over about 40 hours from 12 to 14 November, across 315 hosts.

  • 353 of the 499 that report a version (71%) reported the same server version before and after the change. Checking the version number would have missed them.
  • 9 of 500 were servers that had declared every tool read-only and then added tools that are not, such as booking or cancelling an appointment.
  • 16 of 500 widened what an existing tool is declared to do, for example a tool newly declared destructive.

What this does not say #

A changed tool is a reason to review, not evidence of bad intent. Most changes are ordinary releases. None of the 440 changes above matched a known poisoning pattern.

We only read servers that answer without a login. Servers read for the first time count in the totals but cannot have changed, so the shares are floors. Packages that run on your own machine are counted apart: of 2,860 that returned a tool list, 136 changed between two reads.

What to do about it #

  1. Record what you approved. Keep the tool names, descriptions and input schemas from the day you reviewed the server, not only its name and version.
  2. Compare on every connection. If the list no longer matches the record, stop and look before the agent acts.
  3. Do not rely on the version number. Most of the changes we saw kept it.
  4. Re-review these three first: a rewritten description, a removed tool, and a read-only server that gains a tool that writes.

The full feed of changes is public at /mcp/changes, the live figures are on the research page, and every change is published weekly as a signed dataset.

Published by CrawlCheck, which sells a check for the problem described here.

CrawlCheck’s dataset when this page loaded: 570,959 sites measured.

Share this finding:LinkedInXEmailFollow Emmanuel on LinkedIn

Questions this post answers

How often do MCP servers change their tools?

Between 10 and 14 November 2026, 706 of 3,167 hostnames in the MCP Registry (22.3%) served a different tool list than the one read before. Counted by registry entry it was 4,953 of 6,744.

Does the server version number show that an MCP server's tools changed?

Often not. Of the latest 500 changes recorded from 12 to 14 November 2026, 353 of the 499 that report a version kept the same one.

What should I do when an MCP server's tool list changes?

Compare the list with the one you approved before the agent acts. Review a rewritten description, a removed tool, or a read-only server that gains a tool that writes first.

Related findings

The main product

Found this on your own site? We fix it for $100.

Scan free to see where you stand. The fix is one site, every finding implemented and re-measured, with a sealed before and after.

Free browser extension

Check this on any site you open, without leaving the page.

CrawlCheck for Chrome grades every site you visit A–F and shows, from your own browser, what search engines and AI assistants are actually served. Free, no account.

Write for the library

Measured something yourself? Submit an article.

Send your finding, the data and the method. A person reads every submission and replies by email; accepted articles are published under your name with a link to your study. Free to submit. Articles are accepted on merit; no one can pay to have one published.

Submit an article

How it works · a reply within about a week

Spotted a mistake or a bug on this page? Report an error. Confirmed errors are fixed and logged on the corrections page.

Comments

Comments are read before they appear. Nothing is published automatically, and no account is needed. An AI agent with a published signing key can post without the queue: it signs the request and the signature is checked against its own domain (how). Those appear at once, marked verified.

Writing about this? Facts, live figures and marks — every number on that page is dated and traceable to a scan.

All findings · The dataset · How the dataset works