Library · AI agents and MCP
AI agents and MCP
In one sentence
Findings are short case studies written from real measurements — a robots.txt that answered 200 with a challenge page, a site whose GPTBot traffic came from one address wearing seven crawler names, a homepage that was two percent readable text, a contact email that bounced for one published character — each with the bytes that showed it; a business found with a defect is named only after its owner has been contacted.
What an AI agent can see, use and verify on a site, and the MCP servers agents connect to.
2026-11-02 · Research · Exclusive3 min
AI agents cannot refuse what they cannot see: 27 traps
A revoked credential and an out-of-date fact look exactly like a valid one on the page. So the models walked straight through them.
Read the measurements →
2026-11-02 · Research · Exclusive5 min
MCP Registry census: what 25,983 endpoints answered
We sent the MCP handshake to every remote endpoint the official MCP Registry lists, read the tool lists that came back, and read every flagged description by hand. Most of it is clean. The exceptions matter because an agent connects to them without a person reading a word.
Read the measurements →
2026-11-026 min
Agent readiness checkers: Is It Agent Ready? vs CrawlCheck
Cloudflare’s Is It Agent Ready? and CrawlCheck compared: which agent standards each checks, declared versus working, why a score should skip checks that do not apply, and where each falls short.
ANSWER_ENGINE_REFUSED6.1%ROBOTS_RULES_SHADOWED13%AI_OPTOUT_SET2.9%
Read the measurements →
2026-11-026 min
What is the agentic web? What it means for your website
AI systems now act on websites, not only read them. Which agents actually arrived on our own site, which followed the rules, and the four things you can check this week.
Read the measurements →
2026-11-023 min
How to check an MCP server is safe before an agent connects
An agent that connects to an MCP server reads every tool description as guidance, with no person in between. Five checks, two curl commands and the red flags we found in 25,983 registry endpoints.
INSTRUCTION_CONFLICT0.4%
Read the measurements →
2026-11-024 min
Can an AI agent use your contact form? A two-minute test
When a person asks an assistant to request a quote, the task ends at your form. Five checks in an ordinary browser tell you whether an agent can finish it, and what to change when it cannot.
Read the measurements →
2026-11-02 · Research · Exclusive2 min
MCP reliability, October 2026: 39,484 handshakes measured
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.
Read the measurements →
2026-11-026 min
How an AI agent should verify a website before it acts
In one benchmark run, four AI models caught every lookalike domain on their own and missed the dangers a page cannot show. What an agent should check before it reads, cites, connects or pays, and what counts as evidence.
Read the measurements →
2026-11-02 · Research · Exclusive4 min
MCP servers that ask for a login: what 1,366 request first
We read the login requirements of every MCP server in the official registry that asks for one, without signing in. 44% had at least one finding: most often a default request for every scope the server offers, or no metadata telling a client what to ask for.
Read the measurements →
2026-11-02 · Research · Exclusive4 min
Do AI agents stop before acting on a lookalike? We tested 4
27 tasks, half of them traps: lookalike domains, poisoned tool descriptions, revoked receipts and stale answers. Four models decided alone, then again with a signed answer about the domain. The traps they missed alone were the ones a page cannot reveal.
Read the measurements →