MCP (Model Context Protocol)

What is MCP?

Before MCP, connecting an AI assistant to your CRM, your document store or your database meant a separate integration for every pairing of model and tool. MCP fixes that with a shared language. It’s often compared to a USB port: the tool maker builds one standard connector and any compatible app can plug into it.

An MCP server can offer the model three things: tools (actions such as “create contact” or “look up order”), resources (data to read, such as documents or records) and reusable prompt templates.

How it works

MCP uses a client-server design:

1. Host. The AI application you use: a desktop assistant, a code editor or your own AI agent.
2. MCP client. It lives inside the host and holds the connection to each server.
3. MCP server. It exposes the tools and data of one particular service. It can run on your machine or remotely over HTTP.

On connecting, the client asks the server what it offers. The server replies with its list of tools, their descriptions and the parameters they accept. The model then decides when to use them, through function calling, and the client makes the call. Messages travel as JSON-RPC.

To keep neighbouring terms apart:

Term	What it is
API	A service’s front door: you ask, it answers.
Webhook	The service tells you when something happens.
Function calling	The model decides to call a function, and with what arguments.
MCP	The standard for describing those functions and plugging them into any assistant.

Example

An accountancy practice in Glasgow keeps client files in a shared drive and tracks clients in a CRM. It installs two MCP servers, one per tool, in the AI assistant its team uses. Now someone can ask “which documents is Ruiz Builders still missing for the quarter?” and the assistant checks the CRM, scans the folder and replies with the list. Nobody wrote an integration between the assistant and either tool.

How it affects your business

MCP matters most if your team uses AI assistants every day or if you commission agent development: it cuts the work of connecting each tool and lets you switch assistants without rebuilding integrations. The sensitive part is permissions: an MCP server gives the model real access to your data, so check what it can read and what it can change.

If you build on the UTIO AI API, its capabilities (transcription, text to speech, text, OCR, embeddings, CRM and channels) are ordinary HTTP operations you can wrap as tools in your own server. Every access token carries its own spending cap and rate limit, a useful brake when the caller is a model.