Skip to content
verify mcp Beta VerifyMCP is currently in beta. If you notice any issues, email [email protected] and we’ll put it right.
MCP protocol Scoring Remote servers Packages Building

What is the Model Context Protocol?

The Model Context Protocol (MCP) is an open standard that lets AI applications (like Claude, IDE assistants, and agent frameworks) connect to external tools and data through a single, consistent interface. Instead of every app building a bespoke integration for every service, a service exposes one MCP server and any MCP-compatible client can use it.

If USB-C gave laptops one port for many peripherals, MCP gives AI applications one protocol for many tools.

Clients and servers

MCP has two roles:

  • An MCP host / client is the AI application (e.g. Claude Desktop, Claude Code, an IDE plugin). It starts up, discovers the tools a server offers, and calls them on the model’s behalf.
  • An MCP server exposes capabilities: tools (functions the model can call), resources (data it can read), and prompts (reusable templates). A “GitHub MCP server”, for instance, exposes tools to open issues, read files, and search code.

The client and server speak JSON-RPC and agree on a protocol version before doing anything else. Up to and including 2025-11-25 that agreement happened once, in an initial handshake, and held for the rest of the session; since 2026-07-28 the protocol is stateless, so every request declares its own version and a server publishes what it supports through server/discover. The MCP specification defines two standard transports for that exchange: stdio for a locally-run server and Streamable HTTP for a remote endpoint.

How servers are delivered: packaged vs remote

This distinction is central to trust, and VerifyMCP scores the two on separate rubrics.

  • Packaged servers ship as code you download and run locally: an npm package, a PyPI package, a NuGet package, a container image (OCI), or an MCP bundle (MCPB). Because you execute this code on your own machine, the risks are supply-chain risks: malware, vulnerable dependencies, install scripts.
  • Remote servers are hosted endpoints you connect to over the network (typically Streamable HTTP; the older HTTP+SSE transport is deprecated). You don’t run their code, so the risks are endpoint risks: weak TLS, missing authentication, an unverified operator.

Why trust is hard here

MCP servers are powerful by design: they run code or expose tools that an autonomous agent can invoke without a human in the loop. That same power is the risk surface:

  • A packaged server is a dependency with execution rights on your machine.
  • A remote server is a third party you’re handing context and tool-calls to.

The ecosystem is also growing far faster than anyone can vet by hand, and the official registry lists servers without judging their quality or safety. That’s the gap VerifyMCP exists to fill: an independent, transparent score so you can tell, at a glance, which servers have earned trust.

Common questions

What’s the difference between a packaged and a remote MCP server? A packaged server ships as code you download and run yourself (an npm, PyPI, NuGet package, an OCI image, or an MCPB bundle), so its risks are supply-chain risks. A remote server is a hosted endpoint you connect to over the network without running its code, so its risks are endpoint risks (weak TLS, missing authentication, an unverified operator). VerifyMCP scores the two on separate rubrics.

How does a client connect to an MCP server? The client starts the server (for a local package) or opens a network connection (for a remote endpoint), then performs an initialisation handshake in which the two negotiate a protocol version and exchange capabilities. After that the client discovers the server’s tools, resources and prompts, and calls them on the model’s behalf. Everything travels as JSON-RPC messages.

Do I need a different MCP server for each AI client? No. That’s the point of a shared standard: one MCP server works with any MCP-compatible client (Claude Desktop, Claude Code, IDE assistants, agent frameworks), the same way one USB-C peripheral works across many laptops.

Who created MCP, and is it tied to one AI vendor? MCP is an open standard, published at modelcontextprotocol.io, that any application or service can implement. It is not locked to a single AI provider. (VerifyMCP is an independent project and is not affiliated with Anthropic or any server vendor.)

Next steps

Written by Stuart Blackler · Last reviewed 30 June 2026.