Protocol

MCP (Model Context Protocol)

Why MCP Is Needed

Today, AI calling tools has different solutions here and there — every AI product defines its own tool protocol, and connecting to a database requires writing adapter code separately every time.

The result?

  • The same tool needs to be integrated once by this app, and again by that app
  • Tool descriptions, permission models, and result formats are all incompatible
  • Ecosystem reuse is very difficult

MCP solves this layer of fragmentation.


What Is MCP

One-line definition: MCP (Model Context Protocol) is an open protocol that enables AI applications to connect to external capabilities in a unified way.

Analogy: Like the USB protocol — with a unified interface, mice, keyboards, and hard drives from any manufacturer work when plugged in. MCP is the "USB interface" for the AI world.

With it:

  • Tool providers implement once, and any MCP-enabled client can reuse
  • No need for every AI product to build its own private integration layer

This relates to Tool Calling as: Tool Calling is "the form in which models initiate actions," while MCP is "the standard for tool exposure and integration" — MCP enables tools to be called in a unified way via Tool Calling.


MCP's Three Core Primitives

PrimitiveRoleAnalogy
ToolsOperations the model can callA set of executable functions
ResourcesData the model or application can readRead-only files or data sources
PromptsReusable prompt templatesPre-built workflow entry points

How to Do It: When to Use MCP

Scenarios suited for MCP:

  • A tool wants to be reused by multiple AI clients
  • Needing unified management of tool discovery, permissions, and calling methods
  • Teams building Agents, IDEs, or internal assistant platforms that need a pluggable capability ecosystem

When not to use MCP:

  • Temporarily connecting one internal API within a single project — writing Tool Calling directly is often simpler

Common misconceptions:

  • MCP is not a model's own capability — it's a protocol between AI applications and external capabilities
  • MCP doesn't equal Tool Calling: Tool Calling is the form of model-initiated actions, while MCP is the standard for tool exposure and integration
  • Using MCP doesn't mean it's secure: permission control, sandbox isolation, and audit logging still need to be handled by the application side

Remember this: MCP is the "standard socket" for AI tool integration — implement once by tool providers, usable by multiple AI clients.

Related terms: Tool Calling · Function Schema