Skip to content
RankX AI

Cookie settings

Choose which cookies we can use. You can change this at any time from “Cookie settings” in the footer.

Essential

Security, bot protection and remembering your choice. The site needs them to run, and they do not identify you. Always on.

Google Analytics cookies that tell us which pages are read and which links are followed, so we can improve them. Off until you switch it on.

Tells advertisers when their advert led to your visit. Includes the Meta (Facebook) pixel and the LinkedIn Insight Tag, which set cookies on this domain (LinkedIn also on linkedin.com) and can be told if you later start a trial.

More in the cookie policy and the privacy policy.

AI Agents and MCP

MCP Architecture, Explained Without the Jargon

MCP architecture in plain language: hosts, clients and servers, the two layers, what tools/list and tools/call do, and the 2026 stateless revision.

Published Last updated 7 minute read
Process diagram: Host / client, MCP server, Tool.
On this page 8 sections
  1. What is the Model Context Protocol (MCP)?
  2. Who talks to whom in the MCP architecture?
  3. What are the layers of the MCP architecture?
  4. How does MCP work when an assistant uses a tool?
  5. MCP vs RAG, APIs and function calling: how do they compare?
  6. What changed in the 2026-07-28 revision?
  7. Why does MCP matter for SEO and marketing work?
  8. Follow one catalogue question through the tool call

Short answer: MCP architecture is a client-server model: an AI application (the MCP host) runs one MCP client per connection, and each client talks to one MCP server that exposes tools, resources and prompts. The client lists the server's tools, the model picks one, the client calls it, and the result comes back as data, all as JSON-RPC 2.0 messages.

What is the Model Context Protocol (MCP)?

The Model Context Protocol (MCP) is an open standard for connecting AI applications to external systems: your analytics, your CMS, your visibility data, your database. MCP follows a client-server model, where an AI application embeds an MCP client, an external service runs an MCP server, and the protocol defines how client and server communicate.

Anthropic open-sourced MCP in November 2024, and the specification says it takes some inspiration from the Language Server Protocol, which solved the same many-to-many problem for code editors. Since 9 December 2025, MCP has been governed by the Agentic AI Foundation under the Linux Foundation.

Every message is JSON-RPC 2.0: a method name, parameters, and an id that matches each response to its request. It is a small, well-established format, which is part of why the protocol spread quickly.

Why does MCP architecture matter for AI systems?

MCP makes integration one-to-many instead of pairwise. Before MCP, every AI assistant needed custom code for every service, and LLM applications could only reach the external tools and external data sources their vendor had integrated.

MCP provides one way to connect AI systems to those services: a service publishes one server and every MCP client can use it, whether that client is Claude, ChatGPT, Cursor or a script. That is why connecting RankX AI to Claude is a paste-one-URL job rather than a development project.

Who talks to whom in the MCP architecture?

The MCP architecture has three roles: the host, the client and the server. The MCP host is the AI application you use, such as Claude Desktop, ChatGPT, an editor or another development environment.

How it works
Keep the three MCP roles distinct
  1. Host

    The application the person is using.

  2. Client

    The connection managed by that host.

  3. Server

    The service exposing the tools and resources.

The roles explain the connection. They are not three separate user accounts.

Inside the host live MCP clients, one per connection. The specification says a single MCP client has a one-to-one relationship with a particular server, while one host can run many clients at once.

The MCP server is the service side, exposing what an external system can do. RankX AI's server, for example, exposes tools across visibility, rank tracking, audits and publishing.

What do the MCP host, MCP client and MCP server each do?

Each role has one job, and the table sets them side by side.

Role

What it is

Example

MCP host

The AI application a person uses

Claude Desktop, ChatGPT, Cursor

MCP client

One connection inside the host, to one server

The host's link to a single server

MCP server

The service that exposes tools, resources and prompts

An analytics, CMS or SEO platform

What do MCP servers expose?

MCP servers expose three kinds of thing, called primitives. Tools are functions the model can call, and they carry most of the weight. Resources are data the client can read from a file, document or other data source, and prompts are reusable templates a user can invoke.

Clients have their own feature, elicitation, which lets a server ask the user for more information. Most servers you will meet in SEO work, RankX AI's included, are tool servers.

What are the layers of the MCP architecture?

The MCP architecture has two layers: a data layer, the protocol layer that defines what messages mean, and a transport layer that defines how they travel. Each layer can change without the other, and MCP supports both local and remote servers through it.

The data layer uses JSON-RPC 2.0 for capability and version discovery and for the tools, resources and prompts primitives. In the current revision, a client discovers what a server offers with a server/discover request.

The transport layer carries the messages and handles authorization. The specification defines two standard transports.

Transport

How it works

Where the server runs

stdio

The client starts the server as a subprocess and they exchange newline-delimited JSON-RPC over standard input and output

On your own machine, as a local MCP server

Streamable HTTP

The client sends each message as an HTTP POST to one endpoint, and the server replies with JSON or a request-scoped SSE stream

At one web address, as a remote server

The older two-endpoint HTTP+SSE transport was replaced in the 2025-03-26 revision and is formally deprecated in the current one. The same revision introduced an authorization framework based on OAuth 2.1 with PKCE, which remote servers use when they protect account data.

Because the same data layer runs over either transport, a local MCP server and a hosted one behave identically in conversation. Remote versus local servers covers the hosting choice in its own article.

How does MCP work when an assistant uses a tool?

An assistant uses a tool in a loop: list the tools, pick one, call it, read the result. These five steps are the whole mechanism.

How it works
What one tool call does
  1. Request

    The assistant selects a tool with arguments.

  2. Execution

    The connected server performs that operation.

  3. Response

    The result returns to the assistant for interpretation.

Tool access does not remove the need for permissions and result verification.

  1. Discover the tools

    The client asks the server what it can do with tools/list, and gets each tool's name, description and JSON Schema for its inputs.

  2. Show them to the model

    The descriptions go into the model's context, because an LLM chooses tools by reading about them.

  3. Call the tool

    The model picks a relevant tool, and the client sends tools/call with its name and arguments.

  4. Do the work

    The server runs the tool and returns content such as text or structured data, or a result marked as an error.

  5. Answer from the result

    The result goes back to the model as context, and the model writes its answer from it.

The second step is the quietly clever part, because LLMs pick tools from their descriptions. The tool list is prompt material, so the server explains itself to the assistant in the same language it would use to explain itself to you.

The specification says there should always be a human in the loop with the ability to deny tool invocations. A tool failure comes back as a result with an error flag rather than a dead connection, so the model can read what went wrong and try again.

When you ask Claude which prompts your brand is invisible on and it answers with your own tracked data, it ran that loop against get_prompt_visibility, a tool it discovered when you connected the server.

From RankX AISee which operations the connection supports.A suggested next step is different from applying a product or theme edit.Read the MCP tool reference

MCP vs RAG, APIs and function calling: how do they compare?

MCP, RAG and function calling all bring outside knowledge to large language models, and they sit at different layers. Function calling is how the model picks a tool, MCP is how the tool is discovered and called, and RAG is one of many things a tool can do.

Approach

What it is

Layer

Function calling

A model capability: the vendor-specific way an LLM emits a structured request to run a function

The model

MCP

An open protocol for discovering and calling tools on external systems, whatever AI model the host runs

Between the application and the service

RAG

Retrieval augmented generation: fetch relevant context, often real-time data the model was never trained on, and hand it over

The application

API

A service's own interface, written for developers who read its documentation

The service

Is it MCP versus RAG?

MCP versus RAG is a false choice. An MCP tool that answers a query over your data is retrieval, and MCP gives that retrieval a standard doorway.

MCP does not replace APIs either. An MCP server usually sits in front of an existing API and turns it into self-describing tools, so the question worth asking a vendor is what its server actually exposes.

What changed in the 2026-07-28 revision?

The 2026-07-28 revision, the current MCP specification, makes the protocol core stateless. The release post says any request can now land on any server instance behind a plain round-robin load balancer, so teams can deploy hosted servers like any other web service.

David Soria Parra, MCP's lead maintainer and co-inventor, wrote that the new release is MCP's most important since remote MCP first launched over a year ago. The main changes:

  • The initialize handshake and protocol-level sessions are removed.
  • Servers answer a new server/discover request for capability and version discovery.
  • Roots, sampling and logging are deprecated, with a migration window of at least twelve months.
  • When a server needs more input, it says so in its result and the client retries with the answer, a pattern the changelog calls Multi Round-Trip Requests.

Clients and servers adopt a revision at their own pace, so check which protocol versions a server supports before relying on a 2026-07-28 feature. The mental model does not change: servers describe tools, models pick them, clients call them, and results come back as data.

Why does MCP matter for SEO and marketing work?

MCP enables AI agents and other agentic systems to interact with external systems through one standard doorway, and pre-built MCP servers mean most teams never implement MCP themselves. What matters is what becomes askable once your tools speak it.

Rank tracking, AI visibility, audits and Search Console data join one conversational workflow you can drive in a sentence. Running a full visibility audit from inside Claude shows what that looks like end to end.

One caveat belongs in every explainer: tool descriptions and tool output are untrusted text. Hidden instructions inside them have been shown to steer agents, as in the WhatsApp tool-poisoning case Invariant Labs disclosed in April 2025.

RankX AI's MCP documentation states the rule directly: tool output is data, not instructions. The SEO MCP servers worth connecting are the practical next read.

Follow one catalogue question through the tool call

Suppose a merchant asks which important category pages lost Google clicks. The assistant first needs the correct authorised project, then a tool that can read the relevant report with explicit dates.

The response should carry the page and query evidence the assistant uses to explain the finding. A suggested next step is different from applying a product or theme edit, so check the source report and make any supported write a separate approved operation.

If the tool returns an expired connection or a missing field, the assistant should report that limit instead of inventing a result. MCP standardises the conversation with the service; it does not make incomplete data complete or replace the merchant's judgement.

Next, see the setup instructions or review an agent audit workflow.

Questions about AI Agents and MCP

Is MCP an Anthropic product?

Not any more. Anthropic created MCP and open-sourced it in November 2024, and on 9 December 2025 donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. In 2026 it is an open standard with community governance.

Does ChatGPT use MCP?

Yes. OpenAI's help centre says ChatGPT connects to remote MCP servers, and in the current experience a custom server is added from Plugins. One server can serve every client, so RankX AI's MCP server answers Claude, ChatGPT, Cursor and other MCP clients from the same endpoint.

How is MCP different from an API?

An MCP server usually sits in front of an existing API and turns it into tools a model can discover and call. An API is written for a developer who reads documentation, while an MCP tool carries its own name, description and input schema. The assistant can therefore work out how to use it at call time, without custom integration code.

What is stdio in MCP?

stdio is the local MCP transport: the client starts the server as a subprocess and they exchange newline-delimited JSON-RPC messages over standard input and output. It keeps the server private to your machine, which suits files and desktop software. Hosted servers use Streamable HTTP instead.

How do you know an MCP server can be trusted?

Prefer servers the vendor runs itself, check what each credential is allowed to do, and keep a human able to deny tool calls, as the specification recommends. Treat tool descriptions and tool output as untrusted text, because hidden instructions inside them have been shown to steer agents. A server you install locally can reach whatever your machine can, so audit its code first.

How big is the MCP ecosystem in 2026?

Large and still growing. The July 2026 release post reports close to half a billion SDK downloads a month across the tier-one SDKs, with the TypeScript and Python SDKs each past one billion total downloads. The protocol's December 2025 announcement counted 10,000 active servers, and the PulseMCP directory listed 21,620 on 29 October 2026, many of them forks or experiments.

Sources

  1. Model Context Protocol specification 2026-07-28 (current version) (opens in a new tab) Checked 2026-10-29.
  2. MCP specification 2026-07-28: architecture (opens in a new tab) Checked 2026-10-29.
  3. MCP documentation: architecture overview (opens in a new tab) Checked 2026-10-29.
  4. MCP specification 2026-07-28: transports (opens in a new tab) Checked 2026-10-29.
  5. MCP specification 2026-07-28: tools (opens in a new tab) Checked 2026-10-29.
  6. MCP specification 2026-07-28: changelog (opens in a new tab) Checked 2026-10-29.
  7. MCP specification 2025-03-26: changelog (Streamable HTTP, OAuth 2.1) (opens in a new tab) Checked 2026-10-29.
  8. MCP specification: authorization (current) (opens in a new tab) Checked 2026-10-29.
  9. MCP blog: the 2026-07-28 release (opens in a new tab) Checked 2026-10-29.
  10. MCP joins the Agentic AI Foundation, 9 December 2025 (opens in a new tab) Checked 2026-10-29.
  11. Anthropic: Introducing the Model Context Protocol, November 2024 (opens in a new tab) Checked 2026-10-29.
  12. Anthropic: donating MCP and establishing the Agentic AI Foundation (opens in a new tab) Checked 2026-10-29.
  13. OpenAI Help Center: connecting MCP servers to ChatGPT (opens in a new tab) Checked 2026-10-29.
  14. Docker: MCP horror stories, the WhatsApp data exfiltration issue (Invariant Labs disclosure) (opens in a new tab) Checked 2026-10-29.
  15. PulseMCP server directory (opens in a new tab) Checked 2026-10-29.

Written by

Asif Syed Founder & CEO

Asif Syed is the founder and CEO of RankX AI, the AI search visibility platform. He builds the product and writes here about GEO, AI search measurement and WordPress.

CoversThis article covers the AI Agents and MCP topic, the AI Visibility feature and the AI Readiness Score tool.

Terms usedContext Window, Token (LLM), RAG (Retrieval-Augmented Generation) and Grounding.

Read this page asMarkdown: /blog/how-mcp-works.md.

All articlesEverything RankX AI publishes is listed on the blog index.

Ask an assistantAsk ChatGPT (opens in a new tab), Ask Claude (opens in a new tab) or Ask Perplexity (opens in a new tab).

Preferred sourceIf Google is your front door, you can add RankX AI as a preferred source (opens in a new tab), which asks your own results to surface more of what we publish.

Back to the top

Start here

See which operations the connection supports.

A suggested next step is different from applying a product or theme edit.

7-day free trial. No credit card required. On Shopify, install the app and Shopify bills your plan.