Skip to content
Guide

MCP security scanner: what an automated check can and cannot tell you

The Vouchity team·November 10, 2026·8 min read
Close-up of a computer screen displaying programming code
Photo by Daniil Komov on Pexels

An MCP security scanner is a tool that checks a Model Context Protocol server for risky behaviour before an agent is allowed to use it. There are two different kinds. Code scanners read a server's source and configuration; runtime scanners connect to a live server and test what it exposes. Vouchity does neither of those. Its security signal is an automated heuristic applied to public metadata across the 320 servers it tracks, and this post explains exactly what that heuristic reads, how the numbers are produced, and where it stops.

Key takeaways

  • Vouchity's security check is a scored heuristic over public metadata: authentication, tool permissions, tool-description text and tool count. It is not a manual audit and it does not run the server.
  • Of the 320 servers tracked, 114 (36%) carry a no-auth flag, 4 (1%) request broad host access and 0 (0%) have tool descriptions that read like instructions.
  • The security signal starts at 68, subtracts fixed penalties per finding, and is weighted at 20% of the overall Trust Score alongside maintenance, adoption, transparency and provenance.
  • A serious flag caps the overall score at 84, so a server can never reach grade A while an unresolved risk sits on its card.
  • For code-level checks, pair the heuristic with a static or runtime scanner, and read the tool descriptions yourself with MCP Inspector.

What a security scanner for MCP servers is supposed to find

The risks that matter for MCP are not the same as the ones a web scanner looks for. An MCP server is a set of callable actions handed to a language model, so the questions are about what those actions can reach and what text the model reads before it decides to call them. The OWASP MCP Security Cheat Sheet groups the concerns into areas such as authentication, over-privileged tools and tool poisoning. Tool poisoning is the one that catches people out: the model sees a tool's full description, while a user typically sees only the name, so instruction-like text hidden in the description can steer the agent.

Read our breakdown of the five MCP security risks in the data for the prevalence of each one. This post is about the mechanism: what a scanner can and cannot know from the outside.

The three kinds of check you will see

Most tools that call themselves MCP security scanners fall into one of three groups, and it helps to know which you are buying or running before you trust a result.

  • Static code scannersread a server's source, dependencies and configuration files, and look for patterns such as unvalidated input reaching a shell or a file path. They are strong on code you can see and weak on behaviour that only appears at runtime.
  • Runtime scanners connect to a running server, list its tools, and send test inputs to see what happens. They catch behaviour that code review misses, but they only exercise the paths they are written to try, and they need a server that will accept the connection.
  • Metadata heuristics, which is what Vouchity runs, score what a registry can know without installing anything: whether a remote endpoint declares authentication, what a tool says it needs, how maintained the project looks, and whether the publisher is identifiable. They cover every server at once, but they are inferences from public signals rather than tests of behaviour.

These are complements, not substitutes. A heuristic can tell you a server asks for broad host access; only a code review or a runtime test can tell you whether it actually misuses that access.

How Vouchity's security signal is calculated

The Trust Score has five signals: maintenance, adoption, transparency, security and provenance. Each is scored from 0 to 100 from a concrete, cited public input, and the overall score is a weighted mean of whichever signals are available. The weights are 25% for maintenance, 20% for adoption, 20% for transparency, 20% for security and 15% for provenance. A signal with no public data is left out rather than guessed, which lowers the score's coverage rather than its value.

The security signal begins at a neutral 68 and moves only on four kinds of evidence. Every penalty is a fixed number, so the result can be reproduced by hand from the inputs.

Evidence in the public metadataEffect on the security signalCap
Remote server that declares no authentication−24, and the flag no-auth is raisedApplied once
Remote server that requires authentication+16Applied once
Tools that request broad host access−12 per tool, and the flag broad-permissions is raised−24 in total
Tool descriptions containing instruction-like text−15 per description, and the flag injection-risk is raised−30 in total
More than 40 tools (large attack surface)−8Applied once

The table is the whole rule set for this signal. Nothing in it reads source code, runs the server or looks at the contents of a tool's output.

A worked example you can check yourself

Take a hypothetical remote server. It publishes no authentication requirement, it has two tools that ask for broad host access, its tool descriptions contain nothing unusual, and it exposes twelve tools. Its security signal is worked out as follows:

  1. Start at 68.
  2. No authentication on a remote server: 68 − 24 = 44. The no-auth flag is raised.
  3. Two tools with broad host access: 2 × 12 = 24, which is within the cap of 24. 44 − 24 = 20.
  4. No injection hits and only twelve tools, so no further change. The security signal is 20.

Now the other four signals for the same server. Suppose it was last updated about 80 days ago, which scores 88 on the recency scale. It has 300 GitHub stars, which on a log scale up to 5,000 stars scores about 67. It ships under an MIT licence with a public repository and versioned releases, which scores 87 on transparency. It is active in the official registry but uses an unverified publisher namespace, which scores 34 on provenance. The composite is then:

SignalValueWeightContribution
Maintenance8825%22.0
Adoption6720%13.4
Transparency8720%17.4
Security2020%4.0
Provenance3415%5.1
Overall62100%Grade C

The overall score of 62 is a C, even though the server is maintained, licensed and open. The security signal pulls it down, and that is the point: a server with broad permissions and no authentication should not look like a safe default because it is well maintained. Because the server carries the broad-permissions flag, the cap at 84 would apply only if its score were above 84, and here it is not in play.

What the heuristic cannot tell you

The limits matter more than the arithmetic. The security signal is an automated heuristic and it is labelled as one on every card. Four limits are worth stating plainly.

  • It reads descriptions, not behaviour. A tool can describe itself innocently and still do something harmful. The injection check looks for instruction-like wording; it cannot see what a server does with the data it receives.
  • Broad host access is inferred from wording. A tool that says it reads any path is flagged as requesting broad access, but a tool that is scoped in practice and written vaguely may not be.
  • Absence of a flag is not clearance. The heuristic shows what it found in public metadata. A server with no flags may still have a vulnerable dependency or a poor input-validation habit that only a code scan finds.
  • Coverage varies. Where a signal is unavailable it is excluded and the coverage count drops, so a score built on fewer signals deserves less confidence even when its value looks reassuring.

How to combine it with other checks

The sensible workflow is to use each check for what it is good at. Start with the registry to see the Trust Score, the grade and the plain-English flags for a server. Then use a static or runtime scanner if the server will run in a place where code-level risk matters. Finally, read the server's real tool descriptions and schemas in MCP Inspector before you connect it to an agent that can act on your behalf. Our MCP server vetting checklist sets out the order, and the MCP Inspector guide shows how to do the final step by hand.

For prompt-level attacks in particular, read how prompt injection reaches an agent through tool output. Metadata checks catch the crude cases, but the subtle ones depend on what your agent does with the text it receives.

What to do with a flagged result

A flag is a reason to look, not a verdict. If a server carries no-auth on a remote endpoint, check whether its documentation describes an authentication scheme that the registry metadata simply does not record. If it carries broad-permissions, read the tool names and schemas to see whether the access is needed for the job you want it to do. If it carries injection-risk, open the tool description and read it as the model would. Only then decide whether to connect it, connect it with restricted scopes, or leave it out.

Across the registry, 4 servers (1%) carry the broad-permissions flag, and 0 (0%) carry the injection-risk flag. Those numbers are the count of servers the heuristic flagged, not a measure of how many are compromised. Treat them as the list to review first.

Frequently asked questions

What is an MCP security scanner?

An MCP security scanner is a tool that checks a Model Context Protocol server for risky behaviour. Static scanners read the server's code and configuration, runtime scanners connect to a live server and test it, and metadata heuristics score the public signals a registry can see without running the server.

How does Vouchity's security score work?

The security signal starts at 68 and is reduced by fixed penalties: minus 24 for a remote server that declares no authentication, minus 12 per tool requesting broad host access up to 24, minus 15 per tool description with instruction-like text up to 30, and minus 8 for more than 40 tools. It is weighted at 20% of the overall Trust Score.

Does a clean Vouchity security signal mean a server is safe?

No. The signal is an automated heuristic over public metadata, not a manual audit or a test of behaviour. A server with no flags can still contain vulnerable dependencies or input-validation weaknesses that only a code scan or runtime test would find.

Why can a well-maintained MCP server still get a C grade?

The overall score is a weighted mean of five signals. A server that is actively maintained, licensed and open can still score low on security if it is remote with no authentication or requests broad host access, and that security signal pulls the overall score down.

What should I do if a server is flagged?

Treat the flag as a reason to review, not a verdict. Check whether documentation describes authentication the registry did not record, read the tool names and schemas to see whether broad access is needed, read tool descriptions as the model would, and then decide whether to connect the server, connect it with restricted scopes, or leave it out.

Trust Score changes, in your inbox

A weekly digest of newly flagged risks and the biggest Trust Score movers across the MCP registry. No spam, unsubscribe anytime.

Vet before you connect.

Browse every MCP server's Trust Score free. Create an account to watch the servers you depend on and get notified when something changes.