Skip to content
About · Dresden · est. 2025

The MCP ecosystem
needed a referee.

We built MCPAmpel because LLM agents now install whatever the model recommends, and the model recommends whatever has stars. That's not a security model. So we made the boring one — sixteen scanners, fixed weights, one light.

In one sentence

A trust-light for every MCP server you might install — published methodology, open scoring, German skepticism.

We believe MCP needs a TÜV,
not a leaderboard.

A leaderboard rewards popularity. A TÜV — Germany's vehicle inspection — rewards meeting the same boring criteria as everything else. We're firmly in camp two.

01/03

Open methodology, always

Every weight, every rule, every CVSS bucket is published in the public repo. If you can't reproduce a score on your own machine, it doesn't count as a verdict.

PUBLIC · MIT-LICENSED
02/03

Boring math beats clever models

Sixteen sub-scores, fixed weights, one published nonlinearity. No learned model, no per-server tuning. We trust auditable arithmetic over benchmark-tuned vibes.

NO ML · NO HEURISTICS
03/03

The light is for humans

Red, amber, green. A four-year-old understands it. A CTO understands it. Scores are a tool; the light is the verdict — and that's what goes on the README badge.

UNAMBIGUOUS · LEGIBLE
Why we exist

The agent will install whatever you let it.

In April 2026, the median Anthropic-Claude user has 11 MCP servers installed. Most were added by the agent itself, on the user's behalf, to "solve a task". Many were typo-squat copies. Some leaked credentials. A few were actively malicious.

The npm and pip ecosystems took twenty years to develop social and tooling infrastructure for trust — registries, signing, CVE feeds, security advisories, audit programs. The MCP ecosystem is six months old and growing exponentially. It will not survive its first major supply-chain incident without infrastructure to point at.

MCPAmpel is one piece of that infrastructure. Not the only piece, not the most important — but a piece. A trust-light at the moment of install. A second opinion before the agent runs npm install with your credentials in scope.

04 — Numbers

Where we are so far.

14k+
MCP servers in our index, scanned at least once
2.1k
Servers actively monitored on customer watchlists
135k
Findings catalogued and de-duplicated across the ecosystem
100%
Methodology, rules, and engine versions — public on GitHub
05 — Founder

One person in a Dresden Hinterhof.

N

Nikita Frikh-Khar

IT Sysadmin & Security Researcher, Dresden

HackTheBox Elite Hacker, ranked #16 in Germany. 2x top-5 solo CTF finisher. Reported vulnerabilities to NASA, John Deere, and X/xAI.

Runs IT for two companies. Built their security stack from scratch and blocked real attacks. Built MCPAmpel because AI agents run with real permissions and nobody was checking.

06 — Timeline

From a side-project to a standard.

Sep 2025
v0.1 — three engines, one weekend. Trivy + TruffleHog + a regex for tool-shadowing. Posted on Hacker News, accidentally went front-page.
Nov 2025
First red-light catch. Flagged a popular MCP server with a leaked production AWS key two days before the maintainer noticed.
Jan 2026
v1.0 — sixteen engines, published methodology, NIS2 mapping. Picked up by three EU-regulated banks for their MCP procurement workflows.
Mar 2026
OWASP MCP Top 10 alignment. Co-authored sections 4 and 7. Added the tool-shadowing engine that catches them.
Apr 2026
You are here. 14k servers indexed. 2.1k on monitoring. Public launch.

Want to talk methodology?

We answer every email. Especially the skeptical ones.

[email protected] Read the docs