Skip to content

Methodology

Where the Trust Score comes from

Every Trust Score on Vouchity is computed from real, cited public signals — never modelled, never guessed. Here is exactly how it works.

320

Servers scored

63/100

Registry average

2 months ago

Snapshot refreshed

Five signals, weighted, over what's actually verifiable

A Trust Score is a weighted mean of five 0–100 signals. Each signal is computed only from a concrete, cited public fact — never inferred or defaulted. When a signal genuinely can't be verified for a server (no repo linked, no package found), it is excludedfrom that server's score rather than guessed, and the score's coverage — how many of the five signals were available — is shown alongside it.

Maintenance

25%

How recently the server was actually touched.

Computed from the most recent of a repo push, npm publish, or registry update. Archived repos and publisher-marked deprecated servers are capped low regardless of date; a high open-issue count with no recent activity caps the score further.

Adoption

20%

How widely the server is actually being used.

Log-scaled GitHub stars and weekly npm downloads (whichever is higher, so a CLI-only or GitHub-only project isn't penalised for missing the other). Very low adoption on both axes is a called-out flag, not an automatic failure.

Transparency

20%

Whether you can actually see what you're installing.

A declared license, a public source repository, a homepage, and versioned (tagged) releases each contribute. No license or no public source are both surfaced as explicit risk flags — not just a lower number.

Security

20%

An automated heuristic over public metadata — explicitly not a manual audit.

Starts from a neutral baseline. Declared authentication on a remote server raises it; a remote server with no declared auth lowers it. Tool descriptions implying broad host/filesystem access, instruction-like ('prompt injection') text in metadata, and an unusually large tool count (a wider attack surface) each subtract further.

Provenance

15%

Whether you can tell who actually stands behind it.

An official or reference server maintained by the protocol authors scores highest. Active status in the official registry and a real, namespaced publisher identity (a reverse-DNS handle tied to a domain, e.g. io.github.x or com.acme) each add; an anonymous, unverifiable namespace does not.

Grade bands

The weighted 0–100 score maps to a letter grade. A server with no verifiable signals at all gets ? rather than a fabricated middle score.

GradeScore rangeMeaningIn the registry today
A85–100Trusted4
B70–84Solid84
C55–69Caution147
D40–54Shaky84
F0–39Risky1
?unratedUnrated0

Data sources & refresh cadence

  • Official MCP registry (registry.modelcontextprotocol.io) — a server's identity, namespace, version, transports, package pointers and lifecycle status (active, deprecated or deleted).
  • GitHub REST API — stars, last push date, license, archived status and open issue count for the linked source repository.
  • npm registry — weekly download counts and last publish date for the packaged distribution, where one exists.

Vouchity re-syncs the registry and its enrichment on a regular cadence, and every server page refreshes live via ISR so scores stay current between full re-syncs. The current snapshot was taken 2 months ago.

Honesty, by design

  • Unverifiable signals are excluded, never guessed.A server missing a repo link isn't assumed average on maintenance — that signal is simply left out of its weighted mean.
  • The security signal is a heuristic, not an audit.It reads public metadata for known risk patterns — missing auth on remote servers, broad permission language, instruction-like text, an oversized tool surface. It cannot verify what a server's code actually does at runtime.
  • Risk flags are shown alongside the score, not hidden inside it. Every server page lists the exact flags that pulled its score down and why, in plain English.
A Trust Score is a floor for "this is unlikely to be abandoned, unlicensed or recklessly permissioned" — not a ceiling for "guaranteed safe." Always read the tool list of anything you give real permissions to.

What Vouchity is — and isn't

  • It is a transparent, reproducible scoring layer over public MCP registry, GitHub and npm data.
  • It isn't a manual security audit, a certification authority, or a guarantee of safety.
  • It isn't affiliated with or endorsed by the Model Context Protocol project or Anthropic.

See the score for a server you use.

Search the full registry free, or watch the servers you depend on with Pro.