What is MCP, and why should a website owner care?
Short answer: MCP (Model Context Protocol) is an open standard that lets AI tools such as Claude, Anthropic's AI assistant, and ChatGPT, OpenAI's chatbot, connect to outside systems like your website, calendar, files or customer list. Its own documentation compares it to a USB-C port for AI. For a website owner, it means you can ask an AI tool to read or change things in a connected system, instead of copying and pasting. It is useful but not risk free, because a connected AI can act on real data, so start with read-only access and keep a person in charge.
You keep hearing that an AI tool can connect to your website through MCP and you want to know what that means, what it is good for and how to stay safe.
I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.
MCP is a shared way for AI tools to plug into other software. Before it, every connection between an AI assistant and a business tool had to be built one at a time. With MCP, a tool maker builds one connection point, called an MCP server, and any AI app that speaks MCP can use it. That is why the phrase keeps showing up on website, CMS (content management system), CRM (customer list) and booking tool pages.
This guide explains MCP in plain words, gives business examples you could use, covers the main safety risks, and ends with a first-week plan. It is information for owners, not security advice for a regulated business, so ask a professional where the data is sensitive.
The plain definition
The Model Context Protocol website calls MCP an open-source standard for connecting AI applications to external systems. It says AI apps like Claude or ChatGPT can use it to reach data sources (such as files and databases), tools (such as search engines and calculators) and workflows. Its analogy: MCP is like a USB-C port for AI. One shape of plug, many devices.
Three parts to know:
- The client. The AI app you use, such as Claude, ChatGPT, or Cursor (a code editor with AI built in).
- The server. A small program or hosted service that offers a set of actions, for example "list my pages" or "add a booking".
- The tools. The individual actions the server offers. Each one can read something, change something, or both.
An MCP server does not make an AI smarter. It gives the AI hands. What the AI can do depends on which tools the server offers and which permissions you grant.
Where it came from, and who runs it now
Anthropic, the company that makes Claude, announced MCP on 25 November 2024. On 9 December 2025, the MCP project's own blog said Anthropic was donating MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation (a nonprofit that hosts open-source projects), so that the protocol has vendor-neutral governance. The same post said that after one year MCP had over 97 million monthly SDK downloads (SDK means developer toolkit) and 10,000 active servers. Those are the project's own figures, so treat them as a sign of interest, not an audit.
Why a website owner should care
You may never install an MCP server yourself. You will still run into it, because the tools you pay for are adding it. Three practical reasons:
- Your tools may offer it. Some website and CMS vendors now give you an MCP server, so you can edit or read your site from an AI tool. WordPress.org, for example, lists an official MCP Adapter plugin that exposes WordPress functions to AI agents, with permission checks. For a plainer view of this idea, see what an agentic CMS is.
- Your staff may connect things without telling you. Anyone with a Claude or ChatGPT account can add a custom connector. Claude's help page says custom connectors are available on its Free plan, limited to one there, and on paid plans. You need a simple rule on what may be connected.
- AI tools are becoming a way customers and staff reach your data. Having a clear, safe set of connections is part of running a business with AI in it.
Business examples
These are things MCP connections can do today when the vendor offers a server. Check the vendor for each one.
- Website and content. Ask an AI tool to list pages with missing titles, draft an update, or add a row to a listing. Ploy, an AI website platform, publishes an MCP server with four permission scopes: chat with its agent, read workspace data, write workspace data and publish to production. The publish scope can be withheld, which is a good model for any site. See editing a Ploy site from Claude, ChatGPT and Cursor.
- Calendar and bookings. Ask which slots are free this week, then book one after you approve.
- Customer list (CRM). Ask for every lead from last month with no follow-up, and have it draft the follow-ups for you to review.
- Documents. Ask questions across your price list, procedures and policies, answered from your own files.
- Accounting or orders. Ask for unpaid invoices over 30 days, or the status of an order, in read-only mode.
The common thread is repeat questions and repeat edits, where today you copy data between tabs. For a wider view of what to automate, see business process automation.
How people connect a server (the general steps)
Menus change, so check the current help page of your AI tool. In Claude, the steps on its help page are to open the connectors settings, add a custom connector and paste the server's web address. In ChatGPT, the steps are in the plugins area to add a custom MCP server with a URL and sign-in method. In Cursor, you add the server to a file called mcp.json. The vendor of the tool you are connecting should give you the exact address and steps.
Two limits to know. Claude connects from Anthropic's cloud, so a server must be reachable over the public web. And ChatGPT's custom MCP flow expects read-only search and fetch tools, with write actions needing manual confirmation, per OpenAI's documentation.
The safety side: real risks, in plain words
The MCP tools specification says there should always be a human in the loop with the ability to deny tool use. The AI companies repeat the warning in their own help pages. Claude's page says to connect only to servers you trust and notes that custom connectors are not verified by Anthropic. OpenAI says it does not develop or verify custom MCP servers. Cursor says to install servers only from trusted sources.
The main risks:
- Prompt injection. Text hidden in an email, web page or file that tells the AI to do something you did not ask. If the AI can also write or send, that can cause harm. Claude's help page suggests disabling write tools when you use features that browse or research.
- Too much access. A server with admin rights can delete or change everything. The MCP security guide lists scope minimization (give only the access needed) as a core practice.
- Untrusted servers. A server you did not check could be malicious or sloppy. The MCP security guide also lists risks such as token passthrough and confused deputy attacks, which are technical problems that the server builder has to prevent.
- Silent changes. No log means no way to see what the AI did.
Nobody can call an MCP server "secure" or "approved" in general. You judge each one for the data it can touch.
A copy-ready safety checklist
- Name one person who approves new connections.
- Connect only servers made by the vendor of the tool, or by a developer you can name and contact.
- Start with read-only access. Add write access to one area at a time.
- Keep publishing, payments and deleting behind a human click.
- Test on a copy or a staging site first.
- Check that the vendor shows a log of actions and lets you revoke the connection.
- Never put passwords, card numbers or private customer records into a chat.
- Review the list of connected tools every month and remove what you do not use.
A worked example, with made-up numbers
Hartley Plumbing, a made-up four-person company, thinks about three connections. The owner scores each one for value and risk.
| Connection | Saves (made-up) | Risk | Decision |
|---|---|---|---|
| Calendar, read-only: "what is free Thursday?" | 15 minutes a week | Low | Allow now |
| Customer list, read-only: "who has an open quote?" | 30 minutes a week | Medium (customer data) | Allow after a staff rule |
| Website, write and publish | 20 minutes a month | High (public pages) | Not yet. Test on a copy first |
The order is by risk, not by hype. The lowest-risk connection that pays back goes first. The numbers are made up to show the method.
A test plan with pass and fail rules
- Pick one low-risk connection and give read-only access.
- Ask 10 questions with known answers, such as "how many quotes are open?"
- Pass: at least 9 of 10 answers are correct, the AI says when it does not know, and you can see a log or disconnect in one step.
- Fail: wrong numbers stated with confidence, an action taken that you did not ask for, or no way to revoke access. Remove the connection.
- Only after a pass, test one small write action on a copy, with approval on.
Common misunderstandings
- "MCP is an app." It is a standard, like a plug shape. You use it through an AI app and a server.
- "MCP makes the AI act on its own." The AI still works inside the tool you open, and the specification says a person should be able to deny tool use. How much it does without asking depends on your settings.
- "A connection means the AI has my whole business." It has only what the server offers and what you allow. This is why scoped permissions matter.
- "More connections are better." Each one is another door. Keep only the ones that pay for themselves.
- "MCP helps me rank in AI search." It does not. MCP is about tools you use, not about how AI search tools find your pages.
Another similar name is WebMCP, a separate, browser-side proposal for letting AI agents act on web pages. Do not confuse it with the MCP servers in this guide.
Who can skip MCP for now
- A business that uses its tools only through their normal screens and has no repeat chores worth automating.
- A team that cannot name someone to approve connections.
- Any business handling health, legal or payment data without a professional review of what the server can touch.
- Anyone offered a "secure MCP server" with no named maker, no log and no way to revoke access.
How I'd approach this
I would treat MCP like giving a new employee a key. Start with a key to one room, watch how they use it, and add keys slowly. The value is real when the work is repeated and the data is clean. The risk is real when an AI can change or send things on its own.
Decision rule: connect a tool when it saves you at least an hour a month, can start read-only, and shows a log. Skip it when you cannot name who checks the results. See AI agent guardrails for small business for rules to write down, and what an AI agent is if the word is new. If you want a connected assistant built around your own CRM, calendar and documents, with the limits set up properly and ongoing care, Blur Studio builds that, and you can read what it includes on the business page.
Your first week:
- Day 1: List the tools where you copy data between tabs every week.
- Day 2: Check which of those vendors offer an MCP server, on which plan.
- Day 3: Write the safety checklist above on one page and name the approver.
- Days 4 and 5: Connect one low-risk tool read-only and run the 10-question test.
- Day 6: Score it pass or fail.
- Day 7: Decide on one next step, or stop. Stopping is a fine result.
Related guide: how to update client websites by chatting with Claude or ChatGPT.
Related guide: using Claude and ChatGPT with your website safely.
Questions people ask
What does MCP stand for?
MCP stands for Model Context Protocol. It is an open standard for connecting AI applications to outside systems such as files, databases and business tools. Anthropic announced it on 25 November 2024.
Is MCP only for developers?
Building a server takes a developer, but using one often does not. In Claude and ChatGPT you add a connector by pasting an address and signing in. Each vendor publishes its own steps, so follow the vendor's page for the tool you want to connect.
Is MCP safe for my business data?
It depends on the server and the permissions you give. The MCP specification says a person should be able to deny tool use, and Claude and ChatGPT both warn that custom servers are not verified by them. Start with read-only access and see AI agent guardrails for small business.
Do I need MCP to use AI on my website?
No. You can paste content into an AI tool by hand, or use AI features inside your site builder. MCP saves copy and paste when you do repeat work across tools. For a CMS view, read what an agentic CMS is.
Can MCP help my website show up in AI search?
Not directly. MCP connects tools you control to an AI app you use. It is a different thing from AI search ranking. Google says it needs no special files for its AI features. See SEO vs AEO vs GEO.
What is prompt injection in plain words?
It is hidden text in a page, email or file that tries to give the AI new orders. If the AI can send or change things, it may follow them. Reduce the risk with read-only access, approvals on writes and trusted servers only. Claude's help page also suggests turning off write tools in research features.
Which AI tools support MCP?
The MCP site lists Claude, ChatGPT, Cursor and VS Code (a popular code editor) among supported clients. Plans and limits differ, such as one custom connector on Claude Free and full write access to Ploy's server from ChatGPT only on Business, Enterprise or Edu plans, per Ploy's docs. Check each tool's current help page.
Can I try MCP on a real website server?
Yes, if your vendor offers one. Ploy publishes a server for its own sites with separate permission scopes. See editing a Ploy site from Claude, ChatGPT and Cursor. Use a copy of your site first.
More in this topic: Websites: AI builders, AEO and MCP explained (5 guides) →
Sources
- Model Context Protocol: what is MCP
- Anthropic: introducing the Model Context Protocol (25 November 2024)
- MCP blog: MCP joins the Agentic AI Foundation (9 December 2025)
- MCP specification: tools (human in the loop)
- MCP specification: security best practices
- Claude Help: custom connectors using remote MCP
- OpenAI: MCP and connectors guide
- Cursor: MCP documentation
- Ploy docs: Ploy MCP server and permissions
- WordPress.org: MCP Adapter plugin