What is MCP?

What the heck is MCP?

Model Context Protocol

A shared way for AI apps to connect to tools and data.

MCP is an open protocol, not an AI model.

AI app

MCP server

Your system

The app requests an action or information. The connected system returns a result.

The connector layer behind useful AI.

Sources: MCP introduction and architecture documentation.

How the pieces fit.

AI app

Host + MCP client

Your chat or agent app. The client talks to the server.

MCP server

The connector program

Offers tools, data resources or reusable prompt templates.

Your system

The underlying service

Files, a database or an API. This is where the data or work lives.

A server is software. It can run locally or remotely.

Requests go out. Results come back.

Source: MCP architecture documentation.

Why would you use MCP?

Reuse the connection: Compatible AI apps can use the same MCP server instead of separate adapters.

Discover available tools: The app can list the tools a server offers, including descriptions and required inputs.

Connect to real work: Look up an order, search a document or create a ticket with an authorized tool.

Less custom connector work. Not zero integration work.

Sources: MCP introduction and tools specification. Supported features and permissions vary by app and server.

MCP is not an API key.

API

The interface

A defined way for software to request data or actions from a service.

MCP

The shared protocol

A standard way for AI apps to discover and use a server's tools and context.

API key

A credential

A service may use it to identify a project or authorize requests. Rules vary by provider.

An API key alone does not build an integration.

You can use MCP and an API key together.

Sources: MCP architecture; Google Cloud API keys. Connection method and authentication are separate choices.

When a direct API is enough.

FSR rule of thumb / A known task with a known service

Example: send a new lead to your CRM

Form

Your server

CRM API

API key kept on the server, when supported

You already know the exact action.

A connector or SDK already does the job.

You do not need MCP tool discovery.

Why: fewer layers to run, test and maintain. Some APIs require OAuth or another method, not an API key.

The key grants access. The integration does the work.

FSR guidance, based on the connection architecture. Sources: MCP architecture; Google API keys; Stripe key safety.

When MCP adds value.

FSR rule of thumb / Reusable tools for compatible AI apps

Your AI app benefits from tool discovery.

A trusted MCP server already fits the job.

Several compatible apps need the same tools.

Example: share one order-lookup tool

MCP server

Sales app

Support app

Ops app

Check the actions, permissions and features that each app and server actually support.

AI can use direct APIs, too. MCP is optional.

FSR guidance. Sources: MCP introduction and tools specification. Example assumes apps compatible with the same MCP server.

MCP does not replace security.

Example: a protected remote MCP connection

AI app

OAuth access token issued for this MCP server

MCP server

API key or OAuth token for the downstream service

Business API

Keep secret keys in server-side secret storage.

Limit access. Confirm risky actions before execution.

A connection is not permission to do everything.

Sources: MCP authorization and tools; Stripe key safety. Validate credentials for each service. Use trusted, reviewed tools.

Pick the right connection.

FSR rule of thumb / Choose the smallest setup that fits

Direct API

A known action.

An existing connector or SDK.

No need for an extra MCP interface.

MCP

A compatible AI app.

Discoverable tools and context.

A reusable server that fits the workflow.

Choose authentication separately.

Use the provider's required API key, OAuth flow or other supported method. Both routes need the right access.

Workflow first. Connection second.

FSR decision guide, informed by the official sources. Compare setup effort, running costs, permissions and maintenance.