All posts

MCP AI Claude Shopify

What Is an MCP Server?

September 15, 2026 Agency AI Team 9 min read

META TITLE: What Is an MCP Server? A Plain Guide for Merchants

META DESCRIPTION: What is an MCP server? Model Context Protocol explained in plain terms, plus the Google, Meta and Shopify servers a Shopify store running ads will meet.

SLUG: what-is-an-mcp-server

TAGS: MCP, Model Context Protocol, AI connectors, Claude, ChatGPT, Shopify ads

An MCP server is a program that gives an AI assistant access to one system, such as your ad account. MCP stands for Model Context Protocol. Its official documentation calls it "an open-source standard for connecting AI applications to external systems."

Connected well, one lets you ask Claude or ChatGPT about live campaign data and get a real answer. Connected carelessly, it hands an outside program a door into your accounts.

This guide covers what an MCP server does and how one works. It also covers how MCP differs from an API, and which servers matter to a Shopify store running ads.

MCP Server Meaning, in Plain Language

Model Context Protocol documentation defines MCP servers as "programs that expose specific capabilities to AI applications through standardized protocol interfaces." In merchant terms, a server is a translator sitting in front of one system. It tells the AI what that system can do, then the AI asks in a language both sides agree on.

For a comparison, the protocol's own site reaches for hardware. "Think of MCP like a USB-C port for AI applications." One cable shape, many devices, no separate adapter for each one.

That standard covers three kinds of connection. Its documentation lists data sources like files and databases, tools like search engines, and workflows like specialized prompts.

What Does MCP Stand For

MCP stands for Model Context Protocol. Anthropic released it on November 25, 2024.

On December 9, 2025, Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation. OpenAI's developer documentation calls it "an open protocol that's becoming the industry standard for extending AI models with additional tools and knowledge." Both Anthropic and OpenAI document MCP as an open standard.

That handover announcement put adoption at over 97 million monthly SDK downloads and 10,000 active servers. A Linux Foundation press release the same day cites "more than 10,000 published MCP servers." That is a published count, not an active one.

Both figures come from one coordinated announcement, so neither is an independent measurement. Treat them as a December 2025 snapshot.

Common MCP Servers and What They Connect To

Common examples include file system servers for documents, database servers for queries, GitHub servers for code and Slack servers for chat. One server, one system, one clear job: that pattern holds across all of them.

Most of these do not ask you to write code. Somebody else builds the server and you point your AI app at it. Our roundup of the best MCP servers for marketing covers the ones built for ads and ecommerce.

How Does an MCP Server Work

The Three Parts: Host, Client, Server

MCP's architecture documentation splits the setup into three roles.

An MCP Host is "the AI application that coordinates and manages one or multiple MCP clients." Each MCP Client "maintains a connection to an MCP server and obtains context from an MCP server for the MCP host to use." An MCP Server is "a program that provides context to MCP clients."

Read that as a merchant. Claude or ChatGPT is the host, and the connector plumbing inside it is the client. The server is the program on the other end. Google, Meta, Shopify or your app vendor runs that last part.

Tools, Resources and Prompts

A server can expose three things, and the difference decides who is in charge. Tools are functions the model can call, such as API calls and database queries. Resources are read-only data sources like file contents or database records. Prompts are reusable templates that structure how the model works.

On control, the documentation is explicit:

Tools are model-controlled, so the AI decides when to use one.
Resources are application-controlled.
Prompts are user-controlled, so you pick them.

Tools are the part that acts on your accounts. Per the documentation, tools "may require user consent prior to execution, helping to ensure users maintain control over actions taken by a model."

Local Servers and Remote Servers

A common assumption is that you have to install something. MCP's documentation says otherwise: "MCP servers can execute locally or remotely."

Transport explains why. Local servers talk over standard input and output between processes on the same machine. Remote servers use Streamable HTTP with standard authentication, and the protocol recommends OAuth for tokens. That is why adding a remote server usually means pasting a URL and signing in.

MCP vs API: What Is Actually New

MCP does not replace APIs. It sits above them as one shared way to reach them. According to the specification, the protocol uses JSON-RPC 2.0 messages between hosts, clients and servers.

In the specification's own comparison, MCP does for AI applications what the Language Server Protocol did for development tools. That earlier protocol standardized language support across editors.

Anthropic's launch post describes the problem in its own words, and this is a vendor claim about its own protocol. "Every new data source requires its own custom implementation, making truly connected systems difficult to scale." MCP is offered as "a universal, open standard for connecting AI systems with data sources, replacing fragmented integrations with a single protocol."

APIs still do the work, and MCP standardizes how an AI app asks for it. Our MCP vs API comparison covers when each one fits.

MCP Server Examples for Shopify Advertisers

Google's Ads MCP Server

Google published an open-source Google Ads MCP server on October 7, 2025. Its Google Ads Developer Blog post calls that initial release read-only. In Google's words, it "can be used for reporting and diagnostics but will not make changes to your account."

Setup is self-hosted rather than one click, so it suits a developer more than a merchant. Our roundup of marketing MCP servers lists its tools and setup steps in full. For Shopping and Search reporting, our guide to Google Ads for Shopify is the better starting point.

Meta's Ads MCP Server

Meta runs its own, hosted for you at mcp.facebook.com/ads. Meta's documentation says the server can "create and edit ads, ad sets, and campaigns." On July 16, 2026, following the April open beta, Meta announced that any app can now connect to it.

Access to ad account tools is "gradually rolling out" through this integration method, Meta adds. So what any given account can reach is still moving.

That write access is the important difference. A read-only server can give you a wrong answer. A write-capable server can make a wrong change to a live campaign. Our guide to automating Facebook ads for Shopify covers which automation tools ask for your approval before acting.

Shopify's Storefront MCP Server

Shopify runs a Storefront MCP server on store domains, split across two endpoints. Its documentation describes the purpose as connecting "to a store's catalog, cart, and policies to help customers shop with that merchant."

Cart and policy tools, get_cart, update_cart and search_shop_policies_and_faqs, sit at /api/mcp. Catalog tools search_catalog, lookup_catalog and get_product sit at /api/ucp/mcp.

Two details matter for merchants. Shopify's docs state that Storefront MCP servers "don't require authentication." They also warn that "some stores may restrict access," and advise testing with your specific store.

The Agency AI Connector

Agency AI runs its connector at s.agencyai.app/mcp. Its page names coverage as the difference from single-platform servers: it "connects Meta, Google, and your Shopify store in one place."

This connector reads and writes. Claude can analyze your data, recommend moves and carry them out from the chat. It can create and edit ads, launch and pause campaigns, shift budgets, schedule actions and pull reports.

A write is proposed in the conversation and runs only once you confirm it. Every change it makes shows up in your Agency AI dashboard, alongside the ones you made there yourself. What the connector reads is limited to what you authorize.

In Claude, Agency AI is listed in the connectors directory. In ChatGPT, it is added as a custom connector on Business, Enterprise or Edu workspaces. Our walkthrough for connecting Claude to Meta and Google Ads covers the setup detail.

How to Add an MCP Server to Claude or ChatGPT

Anthropic's support article says custom connectors using remote MCP are available on Claude, Cowork, and Claude Desktop.

Setup takes a few steps. Go to Customize, then Connectors, then add the connector's remote MCP server URL. On Team and Enterprise, the owner adds it under Organization settings, then Connectors. Everyone else then connects to it from Customize.

Our step-by-step Claude setup guide walks through the same flow. ChatGPT adds a checkpoint of its own. OpenAI's documentation states that ChatGPT "currently requires manual confirmation in any conversation before write actions can be taken."

What to Check Before You Connect a Server

Consent comes first in the specification. "Users must explicitly consent to and understand all data access and operations," and hosts "must obtain explicit user consent before invoking any tool." Approval prompts are the design, not an annoyance to click through.

That same specification is blunt about risk. "Tools represent arbitrary code execution and must be treated with appropriate caution." Tool descriptions "should be considered untrusted, unless obtained from a trusted server."

Three checks cover most of it:

Who built and hosts it. Anthropic advises connecting only to servers "built and hosted by organizations and applications you trust."
Whether you trust the operator. Anthropic warns that "malicious MCP servers may include hidden instructions that try to make Claude perform unintended actions."
Whether approval stays on. OpenAI advises keeping approval enabled "for tools that can modify data or take other consequential actions."

A server that asks before every write is the safer way in. You get the analysis, and nothing changes until you confirm it. Check the numbers against a break-even ROAS calculator before you approve any change.

Frequently Asked Questions

What is MCP in AI?

MCP is the Model Context Protocol, an open standard for connecting AI applications to outside systems. It lets an assistant read data and run actions in another tool without a custom integration for each one. Anthropic released it in November 2024 and donated it to the Agentic AI Foundation in December 2025.

Who maintains MCP now?

Anthropic donated MCP to the Agentic AI Foundation on December 9, 2025. That foundation is a directed fund under the Linux Foundation, and Anthropic co-founded it. Anthropic says its commitment is unchanged and that it will keep maintaining the core infrastructure.

Is an MCP server the same as an API?

No. An MCP server usually sits in front of an API and presents it in one standard shape. It uses JSON-RPC 2.0 messages, and the specification compares its role to the Language Server Protocol.

Do I need to install anything to use an MCP server?

Often not. MCP servers can run locally or remotely, according to the documentation, and remote ones are added with a URL. Anthropic's support article describes adding a remote server under Customize, then Connectors.

Is it safe to connect an MCP server?

That depends on who runs it. Users "must explicitly consent to and understand all data access and operations," says the specification. Anthropic advises connecting only to servers from organizations you trust. OpenAI advises keeping approval on for tools that can change data.

Sources

Checked 31 August 2026 against primary pages. Agency AI connector details were rechecked on 14 September 2026.

Protocol definitions, roles, transports, consent rules and the Language Server Protocol comparison come from modelcontextprotocol.io and the MCP specification. Anthropic's launch post of November 25, 2024 supplies the integration problem. Adoption figures come from the MCP blog and the Linux Foundation press release of December 9, 2025.

Google's read-only status comes from the Google Ads Developer Blog of October 7, 2025. Meta's write capability and July 16, 2026 opening come from Meta's docs and blog. Shopify's docs cover the Storefront endpoints.

Setup steps and trust warnings come from Anthropic's and OpenAI's support docs. Agency AI details come from agencyai.app/mcp and are vendor material.

Free to install. Launch campaigns and pay nothing for your first 30 days.