Edit Your Website From Claude or ChatGPT (MCP)

The Short Answer

You can edit a website from Claude or ChatGPT when the site exposes an MCP server: a small service that gives the AI a defined set of tools, such as "read this page" or "change this headline". You add the server's URL in your assistant's connector settings, sign in, and from then on you can type "change the opening hours on the contact page to 9 to 5" in the same chat you already use for everything else. The assistant calls the tool, and the change lands in the site. What matters is not the chat window but what the tools allow, so most of this guide is about that part.

What MCP Is, in Plain Terms

MCP stands for Model Context Protocol. The official introduction describes it as "an open-source standard for connecting AI applications to external systems" and compares it to a USB-C port: one standard plug, many devices. Claude, ChatGPT, Cursor and VS Code all speak it, so a service that builds one MCP server can be used from all of them.

The architecture overview names three roles:

The host is the AI application you talk to, such as Claude Desktop or Cursor.
The client is the connection the host opens for each server it uses.
The server is the program that offers context and actions.

A server can offer three kinds of things. Tools are actions the AI can run ("file operations, API calls, database queries"). Resources are data it can read. Prompts are reusable templates. For website editing, tools are the part that counts: they define exactly what the AI is allowed to do to your site.

Servers run in two ways. A local server runs on your own machine and talks to the app over standard input and output. A remote server runs on the web and is reached over HTTP; the specification recommends OAuth for signing in. Website editing services are almost always remote servers, because the site's code and hosting do not live on your laptop.

What an MCP Server for a Website Actually Does

When you ask Claude to change a headline on your site, the assistant does not see your website the way a browser does. It asks the server what tools exist, reads their descriptions, picks one, and calls it with arguments. A typical sequence for a text edit looks like this:

The assistant calls a read tool to get a map of the site: pages, sections, which file holds which text.
It reads the relevant file or record to find the exact wording.
It calls an edit tool with the old text and the new text.
The server decides what happens next: write the change, ask for confirmation, or route it to someone for review.

Step 4 is where website MCP servers differ most. A server that hands the AI a raw "write any file" tool turns every chat into a deploy with no safety net. A server built for content editing keeps the decisions on its side: which files are editable, which changes need a human, and what gets recorded.

How to Connect Each Assistant

The details vary by app and change often, so check each vendor's current docs. The general shape:

Claude (claude.ai and Claude Desktop)

Anthropic calls remote MCP servers custom connectors. You add one in the connectors section of Claude's settings by entering the server URL, then choose how to sign in. Free plans are limited to one custom connector. In a chat, you turn connectors on or off from the "+" menu. Anthropic's own warning is worth repeating: custom connectors are not verified by Anthropic, so only connect servers run by organisations you trust.

ChatGPT

OpenAI's MCP documentation describes adding a custom MCP server with its URL and authentication details, then testing it in a chat. ChatGPT asks for manual confirmation before write actions run, which is a useful default for anything that edits a live site.

Cursor and Claude Code

Developer tools take the same URL in a config file. Cursor reads remote servers from .cursor/mcp.json in a project or ~/.cursor/mcp.json globally, supports OAuth sign-in, and, per Cursor's MCP docs, asks for approval before using MCP tools by default. A minimal entry looks like this:

{
  "mcpServers": {
    "my-site": {
      "url": "https://mcp.example.com"
    }
  }
}

For a developer, this is handy when content edits come up in the middle of coding: you do not need to switch to another dashboard.

The Guardrails to Insist On

The assistant's own confirmation prompts help, but they protect you from the AI acting without asking, not from the AI asking for the wrong thing. If you are choosing an MCP server for a client site, or building one, these are the questions that matter.

Does it limit what can be edited?

The safest servers know which parts of the site are content and which are code. Changing the text inside a heading is routine. Changing the component around it, the layout, or a config file is not. Look for a server that classifies files or zones and refuses, or escalates, edits outside the content areas. Without that, a polite request like "make the hero section shorter" can end up rewriting markup.

Who reviews risky changes?

Some edits are fine to publish straight away: a typo, a phone number, a price in a pricing table. Others deserve a second pair of eyes. In a Git-based site, the natural way to get that review is a pull request: the developer sees the diff, approves or rejects it, and nothing reaches production until they do.

Is there a record and a way back?

Every change should be traceable to a person and a time. On a site that lives in Git, commits give you that for free, along with the option to revert. If the server writes to a database instead, ask how changes are logged and how you undo one.

Does it protect search rankings?

The quickest way to lose traffic with an AI editor is to let it rewrite a page title, H1 or meta description that already ranks. A careful server treats those fields as sensitive and warns before touching them. Google's own guidance on title links shows how much the title and main heading feed into what searchers see.

What does the AI get to see?

The assistant needs the text it is editing. It does not need your deploy keys, database credentials or API tokens. Those should stay on the server side, and you should be able to revoke the assistant's access in one step.

When MCP Is Not the Right Tool

MCP editing is a good fit when the people making changes already live in an AI assistant and the site's content sits in files or a database the server understands. It is a poor fit in a few cases:

Large structured content teams. If a dozen editors publish articles daily with workflows, scheduling and roles, a dedicated headless CMS with an editorial interface is built for that. We compare the two approaches in CMS Chat vs headless CMS.
Visual redesigns. Moving sections, changing a layout or adding a new page type is development work. A chat request can describe it, but a developer should build it.
People who do not use an AI assistant. For a client who never opens Claude or ChatGPT, a chat box inside a site portal is simpler than setting up a connector. Our guide on how clients can edit website text without code covers the other options.

Where CMS Chat Fits

CMS Chat connects to a site's GitHub repository through a GitHub App, detects the framework, and maps the editable zones: text, images, links and prices in TSX, MDX, Markdown, HTML and JSON files. Each zone gets a guardrail: editable, warn, escalate or locked. Routine edits are committed, risky ones become a pull request, and layout or code changes are never guessed at.

The same guarded tools are available over MCP at https://mcp.cmschat.ai. You add that URL as a connector in Claude, ChatGPT, Claude Code, Codex or Cursor, sign in with your CMS Chat account, and choose which site the assistant may edit. Zone rules and escalation live inside the tools, not in the assistant, so they apply whichever app you use. The MCP setup guide has the steps for each client, and Connect AI shows the short version.

CMS Chat is in early access. If you run client sites built in code and want them editable from the assistants your clients already use, request a beta invite or read how it works first.

Frequently Asked Questions

Can ChatGPT or Claude edit my website directly?

Not on their own. They need a connection to the site, and MCP is the standard way to give them one. Without an MCP server, the assistant can only suggest text for you to paste in yourself.

Is it safe to give an AI write access to my site?

It depends on the server, not the assistant. Look for limits on which files can change, review for risky edits, a record of every change, and credentials that stay on the server side. Keep the assistant's confirmation prompts switched on for write actions.

Do I need to be a developer to set this up?

For the person editing, no: adding a connector means pasting a URL and signing in. Someone technical usually sets up the server side once, for example by installing a GitHub App on the site's repository.

What is the difference between an MCP server and a chatbot plugin?

A chatbot plugin usually adds a chat widget to your website for visitors. An MCP server works the other way round: it lets your own AI assistant reach into the site to read and change it.