MCP & Trust Score glossary

Model Context Protocol (MCP)

An open standard that lets AI models call external tools, data sources and services through a uniform interface. An MCP server exposes capabilities; an MCP client (your agent or IDE) discovers and invokes them.

MCP server

A program that publishes a set of tools, resources or prompts over MCP so an AI agent can use them — for example a GitHub server, a Postgres server or a browser-automation server.

MCP client / host

The application an agent runs inside — Claude Desktop, an IDE extension, a custom agent — that connects to MCP servers, lists their tools and calls them on the model's behalf.

Tool

A single callable capability an MCP server exposes, with a name, a description and a typed input schema. The model reads the description to decide when and how to call it, which is exactly why an untrustworthy description is a risk.

Transport

How a client talks to a server. Local servers usually use stdio (spawned as a subprocess); remote servers use streamable-HTTP or SSE over the network. Remote transports widen the attack surface, which the Trust Score accounts for.

Local vs remote server

A local server runs on your own machine and is reached over stdio; a remote server runs on someone else's infrastructure and is reached over HTTP. Remote servers can see your requests and require auth to be safe.

Trust Score

A single 0–100 number Vouchity computes for every server from five weighted, independently verifiable signals. It is a starting point for due diligence, not a manual security audit or a guarantee.

Grade (A–F)

The letter grade the 0–100 Trust Score maps to — A for trusted, C for caution, F for risky. A serious risk flag caps the grade so a well-maintained but insecure server can't score an A on maintenance alone.

Coverage

How many of the five signals could actually be verified for a server, out of five. A server with a repo, npm package and declared auth has higher coverage — and a more meaningful score — than a registry-only entry.

Maintenance signal

Whether the project is actively kept up — recency of the last commit or release, and whether the repository is archived. Stale or abandoned servers score low here.

Adoption signal

How widely the server is used and trusted by others — GitHub stars and npm weekly downloads. Broad adoption means more eyes on the code and faster discovery of problems.

Transparency signal

Whether you can actually inspect what you're installing — a public source repository, a declared license and versioned releases. No license and no source both lower this signal.

Security signal

An automated heuristic over the server's posture — declared authentication, requested permissions, and injection-shaped patterns in tool descriptions. It flags risk; it is not a manual audit.

Provenance signal

Where the server comes from and how it's published — presence in the official MCP registry, a namespaced publisher, and official-vendor status. Unknown provenance lowers confidence.

Risk flag

An explicit, named warning attached to a server — such as no-auth, archived, deprecated, no-license, broad-permissions or injection-risk. Flags are surfaced verbatim so you see exactly what dragged a score down.

No-auth endpoint

A remote server that declares no authentication requirement. Anyone who knows the URL can invoke its tools, which is a meaningful risk for a server that can touch data or systems.

Prompt injection (tool-description)

Malicious instructions hidden in a tool's name or description that try to hijack the model when it reads them. Injection-shaped text is a security red flag because the model treats descriptions as trusted context.

Official MCP registry

The canonical, community-run list of MCP servers at registry.modelcontextprotocol.io. Anyone can publish to it — which is precisely why an independent trust layer on top of it is useful.

GEO (generative engine optimization)

Structuring content and data so AI answer engines can find, understand and cite it accurately — the machine-readable cousin of traditional search optimization.

llms.txt

A plain-text file at a site's root that gives language models a concise, machine-first map of what the site covers and how to use its data. Vouchity publishes one so agents can discover the Trust API and MCP endpoint.

Ready to see the terms in action? Browse every scored server on the registry, read the full scoring methodology, or start with the explainer What is an MCP server?