What Is MCP (Model Context Protocol)? A Complete Guide

MCP (Model Context Protocol) is an open protocol that lets AI assistants connect to external systems in a standard way. Through MCP, Claude, ChatGPT and similar assistants can read records from a database, call an API, or change a budget in your ad account. This guide covers what MCP is, why it exists, how it works technically, and how hosted and local servers differ.

What is MCP (Model Context Protocol)?

MCP stands for Model Context Protocol. Anthropic published it as open source in November 2024, and it became an industry standard quickly enough that applications well outside Anthropic now support it.

The one-sentence definition: MCP is a shared language describing which tools an AI assistant can call and how it should call them. Because that shared language sits between the assistant and the external system, nobody has to write bespoke glue code for every new pairing.

An analogy helps. Just as USB-C lets different devices speak over one cable, MCP lets different AI applications speak to different systems through one interface.

Why was MCP created?

Before MCP, every AI application needed its own integration for every system it wanted to reach. Three assistants and ten systems meant thirty separate integrations, each with its own maintenance burden, failure surface and security review.

MCP broke that multiplication. You write one server for a system, and every application that speaks the protocol can connect to it. The server author does not need to know which client is calling, and the client does not need to know what is inside the server.

How does MCP work?

The architecture has three parts:

1. MCP host (the client): the application the assistant runs in, such as Claude, ChatGPT or Cursor. It takes your request and lets the model decide which tool to call.
2. MCP server: the side that defines the tools and makes the real calls. It tells the client each tool's name, purpose and expected parameters.
3. Data source: the actual system behind the server. A Google Ads account, a database, a file system.

The flow runs like this: you ask for something in plain language, the model picks a suitable tool, the server executes it against the real system, the result goes back to the model, and the model writes you a readable answer. One detail matters more than the rest: the model chooses the tool, but the server decides what is possible. A model cannot invent an operation the server never defined.

How does the MCP protocol work technically?

MCP uses JSON-RPC 2.0 for messaging. Client and server declare their capabilities to each other when the connection opens, so both sides know what the other supports.

The protocol defines three core primitives:

• Tools: actions the model can invoke, such as pulling a report, creating a record or changing a budget.
• Resources: pieces of data the model can read, such as a file, a record or a configuration.
• Prompts: reusable templates the server offers for recurring tasks.

Connections run over two transports. stdio is for local processes, where the server runs as a program on your machine. HTTP based transport connects to a remote server and is what hosted servers use. That difference drives everything in the next section.

Remote MCP vs local MCP

An MCP server runs either on your computer or in the cloud. Both speak the same protocol, but the day-to-day experience is very different.

Local (self-hosted) MCP server

Runs on your own machine. Google's open-source Google Ads MCP server is in this category.

What it requires:

• A terminal environment with Python installed
• Creating a Google Cloud project
• A developer token application, where approval can take weeks
• Setting up OAuth 2.0 credentials by hand
• Owning updates and maintenance yourself

Weak points: setup takes hours or days, the server is down whenever your computer is off, and these servers are usually read-only.

Remote (hosted) MCP server

Runs in the cloud. You paste one URL.

Strong points:

• Setup takes about 60 seconds
• No terminal, Google Cloud project or developer token
• Keeps running when your computer is off
• Updates are automatic, with no maintenance burden
• Supports writes as well as reads
• Handles multiple accounts, including MCC, on one connection

Which AI applications support MCP?

• Claude (Pro, Max, Team, Enterprise) via claude.ai and the Claude desktop app
• ChatGPT via the connectors section
• Developer tools such as Cursor and Claude Code
• Other applications that implement the protocol

This is where a hosted server pays off: the same URL works in every application, so switching tools does not mean reinstalling anything.

Managing ad accounts with MCP

Ad operations is one of the most concrete uses of MCP. Opus Growth is a hosted MCP server for Claude and ChatGPT, and one URL connects five ad platforms: Google Ads, Meta Ads, Microsoft Advertising, TikTok Ads and LinkedIn Ads. The same sign-in also covers Search Console, GA4, YouTube and Tag Manager data.

A sample request: "Find search terms with no conversions in the last 30 days and add the highest spenders as negatives." The assistant pulls the report, ranks the terms, then shows you a preview before writing anything. Nothing changes until you approve it.

On the read side you get performance reports, search-term analysis, keyword quality scores, account audits and change history. On the write side you get campaign, ad group, keyword and ad copy creation, budget and bidding management, plus extensions and targeting.

Setup: get started in 3 steps

1. Add the URL: in Claude, open Connectors, add a custom connector and paste https://mcp.opus-growth.com/mcp.
2. Connect your accounts: pick the accounts you want managed in the authorization window that opens.
3. Manage by chatting: write what you want done in plain language.

Security: how is access limited in MCP?

In MCP the model's reach is bounded by the tools the server defines. If the server never defined an operation, the model cannot perform it. The rest of the security story belongs to whoever wrote the server.

On the Opus Growth side, the safeguards are:

• Every write is shown as a preview first
• Nothing is applied without explicit approval
• A single change can never raise a budget by more than 50 percent
• Authorization keys are stored encrypted
• Every applied change is written to an audit log

Frequently Asked Questions

What does MCP stand for?

MCP stands for Model Context Protocol. Anthropic published it as open source in November 2024.

What does the MCP protocol actually do?

It standardizes how AI assistants talk to external systems. One server written for a system can be used by every assistant that speaks the protocol, which removes the need to build a separate integration for each assistant and system pair.

Does the MCP abbreviation mean anything else?

It does. MCP is also used in medicine for the drug metoclopramide and in electronics for various components. The MCP in this article is the software protocol, Model Context Protocol.

Should I choose a remote or a local MCP server?

If you prefer writing code and running your own infrastructure, a local server fits. If you want to skip setup entirely, keep the server running while your computer is off, and perform writes, a hosted server fits better.

Do I need to know how to code to use MCP?

Not for a hosted server, where you only paste a URL. Running your own server requires terminal familiarity and basic configuration work.

Is MCP secure?

The protocol itself limits the model to the tools a server defines, so a model cannot invent new access on its own. Practical security depends on the server's approval flow and permission boundaries. Prefer a server that previews writes and asks for explicit approval.

Which platforms besides Google Ads can I manage?

Through the Opus Growth connection you can manage Google Ads, Meta Ads, Microsoft Advertising, TikTok Ads and LinkedIn Ads. The same authorization also covers Search Console, GA4, YouTube and Tag Manager.

Is there a free trial?

Yes. You start with a trial that asks for no card, and the credits it gives you never expire.

Trial and pricing

• Free trial: 75 credits that never expire, no credit card required.
• Pro ($49/mo): one ad account, 5,000 credits per month.
• Agency ($99/mo): multiple accounts and MCC management.

Try Opus Growth | See pricing

Related posts: Managing Google Ads with Claude · Managing Google Ads with ChatGPT · Safe AI Ad Management