CrawlCheck

Findings · 2026-11-13 · By

Both our SDKs cut a tool list at 300 with no warning

A fingerprint of an MCP server's tools is only useful if it covers all of them. Ours stopped at 300 and did not say so.

Until 11 November 2026, CrawlCheck's MCP lockfile and both SDKs fingerprinted only the first 300 tools a server listed and did not mark the result as cut. A change to tool 301 or later left the fingerprint unchanged. Now every tool is listed, a list that could not be read in full is marked incomplete, and a check against an incomplete list answers block. Upgrade to @crawlcheck/sdk 1.14.0 or crawlcheck 1.15.0 on PyPI, or later.

Three cases, before and now
Server lists 301 tools
Before
Fingerprint of the first 300, no mark
Now
All 301 listed
Server’s list could not be read to the last page
Before
Fingerprint of what was read, no mark
Now
Marked incomplete
Checking against an incomplete list
Before
Could pass
Now
Answers incomplete, decision block

An MCP server tells an agent which tools it offers. We record a fingerprint of that list so an agent can notice when a server's tools change after it was approved. The point of the fingerprint is that any change shows.

The defect #

Our code read up to 300 tools and built the fingerprint from those. A server with more than 300 got a fingerprint of the first 300, and nothing in the result said the list had been cut. If tool 301 changed, or a new tool was added past that point, the fingerprint stayed the same and a check against it passed.

Most servers list far fewer than 300 tools, which is why this sat unnoticed. That does not make it acceptable: the servers with the most tools are the ones where a silent change is easiest to hide.

Before and after #

CaseBeforeNow
Server lists 301 toolsFingerprint of the first 300, no markAll 301 listed
Server's list could not be read to the last pageFingerprint of what was read, no markMarked incomplete
Checking against an incomplete listCould passAnswers incomplete, decision block

What we changed #

The lockfile now lists every tool. The SDKs read the server's tool list to its last page. Where a list still cannot be read in full, the result says so, and a check against it does not pass. We confirmed it on the packages as published: a 301-tool list comes back with 301 entries.

What to do #

If you use the JavaScript SDK, move to @crawlcheck/sdk 1.14.0 or later. If you use Python, move to crawlcheck 1.15.0 or later. If you pinned a lockfile for a server with more than 300 tools before 11 November 2026, create it again, because the old one does not cover the whole list.

If you build something similar, the rule we took from this: a limit is fine, a silent limit is not. Anything that stops early has to say that it stopped. More on what agents can verify is in our guide to agent checks.

CrawlCheck’s dataset when this page loaded: 570,327 sites measured.

Share this finding:LinkedInXEmailFollow Emmanuel on LinkedIn

Questions this post answers

What was wrong with the CrawlCheck SDKs?

They fingerprinted only the first 300 tools of an MCP server and did not mark the result as cut, so a change past tool 300 went unnoticed.

Which versions fix it?

@crawlcheck/sdk 1.14.0 on npm and crawlcheck 1.15.0 on PyPI, and every later version.

Do I need to recreate my lockfile?

Only if the server lists more than 300 tools and the lockfile was made before 11 November 2026. For smaller servers the fingerprint already covered everything.

What happens now when a tool list cannot be read in full?

The result is marked incomplete, and a check against it answers incomplete with the decision block instead of passing.

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