No, you can't point an AI at a URL and watch it rewrite the page.
That's the version most people arrive with: paste your homepage address into ChatGPT, describe a change, and the live site quietly updates itself. MCP doesn't do that, and never claimed to. Model Context Protocol is an open standard, released by Anthropic in late 2024, that lets an AI assistant connect to a system that has deliberately exposed a set of actions to it. For websites that resolves into two genuinely different things: operating and editing your own site by chatting to an assistant, and making your site usable by other people's AI agents (a related browser standard called WebMCP). Nothing happens by URL alone; the platform on the other end has to speak the protocol. Devani, the free open-source CMS running this blog, is built to speak it natively, which is why you can connect your own ChatGPT or Claude and edit the site by describing what you want.
The assistant (the client)
The AI app you already talk to, ChatGPT or Claude, plays the role MCP calls the client. When you connect it to a website's MCP server, it can see the actions that server offers and call them on your behalf. You bring your existing account; there's no separate model to license.
The server (the bridge)
An MCP server sits in front of a system and publishes what it can do as structured tools: create a page, update this heading, list the blog posts, set a redirect. The assistant reads that menu and picks the right tool with the right inputs. No screen-scraping, no guessing which button to click.
No API keys, no code
Because the assistant is your own logged-in ChatGPT or Claude, you're not pasting API keys or writing integration code. You authorize the connection once, then work in plain language. The protocol carries the request; the server enforces what's allowed. That's the whole point of standardizing it.
Reversible by design
The good MCP-for-editing setups snapshot before they act. On Devani every change the assistant makes is captured, so a rewrite you don't like is one click back. That safety net is what turns "let the AI edit the live site" from a nervous idea into a normal Tuesday.
Same three letters, two different jobs.
Search "MCP for websites" and the results split down the middle. It's worth knowing which one you actually want before you go shopping, because the tools don't overlap.
The first job is the one this article opened on: you, the owner, editing and building your site by talking to an assistant. Your CMS runs an MCP server, your ChatGPT or Claude connects to it, and you manage the site in conversation instead of clicking through an admin panel. This is what Webflow, Wix, WordPress.com, and Devani mean when they talk about MCP. It's a private, authenticated connection between you and your own site.
The second job points outward. WebMCP, a separate browser standard announced in February 2026 and co-authored by Google and Microsoft through the W3C, lets your website hand its own features to a visitor's AI agent as callable tools, through a new navigator.modelContext browser API. Instead of an agent like Gemini in Chrome taking a screenshot of your booking form and guessing where to click, your page tells it directly: here's the "book appointment" action, here are the fields it needs. One is about you commanding your site; the other is about strangers' agents using it. People conflate them constantly, and they solve opposite problems.
One protocol lets you run your site by talking. The other lets your site talk back to somebody else's AI. Don't buy the wrong one.
The platforms that speak MCP, and how deep it goes.
By mid-2026 most serious platforms have shipped something MCP-shaped. What varies wildly is whether the assistant genuinely operates the site or just drives a slice of it.
| Platform | What its MCP does | Where the AI actually works |
|---|---|---|
| WordPress.com / Pressable | Create sites, edit content, audit hosting from ChatGPT, Claude, or Gemini; the Vibe AI plugin adds in-editor building | Best inside the block editor; page-builder stacks like Divi and Elementor fall back to plain HTML |
| Webflow | API-backed server: create elements and styles, manage collections, assets, and custom code | Powerful, but a developer-oriented tool with a real learning curve of its own |
| Wix | Official MCP server compatible with Claude, Cursor, and other MCP clients | Operates the platform through its API rather than a plain-HTML canvas |
| Sitecore, Optimizely, Umbraco | Enterprise CMS MCP servers for editorial and experimentation teams | Admin-facing, aimed at large content operations |
| Devani | MCP-native: connect your own ChatGPT or Claude, edit content and layout, every change snapshotted with one-click rollback | Inside the visual builder, because pages are plain HTML the assistant writes cleanly |
That last column is the whole game. Almost everyone can now say "we have MCP." The honest question is what the assistant can touch once it's connected, and how much setup stands between you and the first edit.
Bolted onto the builder, or built into it.
The difference that matters isn't whether a platform ships an MCP server. It's whether the assistant works inside the thing you build with, or beside it.
Bolt-on MCP
An AI server added to a platform that wasn't designed around it.
- ✓ Drives the platform you may already run, through its own API
- ✓ Genuinely useful for content edits and site management
- × WordPress.com's MCP and the Vibe AI plugin work well in the block editor; on Divi or Elementor the assistant can't reliably produce the builder's shortcodes, so you get AI-written content plus the standard HTML editor, not a builder
- × Developer-oriented servers (Webflow, Wix) often assume tokens, technical setup, or a designer's tool underneath
Devani
MCP-native, so the assistant is the building surface, not an add-on.
- ✓ Connect your own ChatGPT or Claude subscription, no API keys, no copy-paste, no plugin to install first
- ✓ Edits live content and layout, because pages are plain HTML the assistant writes easily and lands in the visual editor
- ✓ A whole site can be built with AI in 10 to 15 minutes and go live in under a minute on the managed free tier, and it stays fast: database-free, served as static files from a CDN, around 100 out of 100 Lighthouse with a 0.7-second desktop LCP
- × Younger and smaller than WordPress or Webflow, with fewer templates and no plugin marketplace; a build that leans on one specific niche plugin should stay where that plugin lives
Be precise about the claim, because everyone is catching up here and the honest line is narrow. Devani isn't the only platform with MCP, and pretending otherwise would age badly. What's defensible is that it's MCP-native and the assistant works hand in hand with the builder from inside it. That's a property of the architecture, not a feature bolted on: a database-free site is just files of clean HTML, which is exactly the kind of thing a language model reads and rewrites without tripping. On platforms where a page is an opaque tangle of proprietary components, the assistant either works through an API layer or gets stuck at the editor's edge. If you already live in WordPress, Webflow, or Wix, their MCP server drives the platform you know, and that's a real reason to use it. If you're choosing fresh and want the assistant to be the way you build, native beats bolted-on.
Making your site legible to other people's agents.
WebMCP is the second meaning, and it's early. If AI agents start doing real shopping, booking, and form-filling on behalf of their owners, this is how your site stops being a screenshot they squint at.
- Two ways to expose tools. The declarative API adds attributes to HTML forms you already have, so a contact or booking form becomes an agent-callable action with very little code. The imperative API uses a JavaScript
registerTool()call for dynamic workflows that need validation or custom logic. - It inverts the guessing. Rather than an agent reverse-engineering your interface from pixels, your page hands it an explicit list of actions and the exact parameters each one takes. Faster, more reliable, and far cheaper in tokens than vision-and-click.
- It's a standard in motion, not a finished one. As of mid-2026 it's a W3C effort in a Chrome origin trial with Gemini in Chrome as the main agent consuming it. Firefox and Safari haven't committed publicly. Treat it as forward-looking infrastructure, not a box to check this week.
- Clean HTML is the head start. Whichever way this shakes out, the sites easiest to make agent-ready are the ones already built from plain, semantic HTML rather than a rendered soup of framework markup. The same thing that makes a site fast and easy for an assistant to edit makes it easier to expose to agents later.
What is MCP for websites, in one sentence?
It's an open standard that lets an AI assistant connect to a website's system and use a defined set of actions, so you can build or edit your site by chatting instead of clicking through an admin panel. The site has to expose an MCP server for any of it to work; you can't just aim an AI at a URL.
What's the difference between MCP and WebMCP?
MCP connects an assistant to a system's tools, and for websites that usually means you editing and managing your own site by talking to ChatGPT or Claude. WebMCP is a separate browser standard that lets your site expose its features to a visitor's AI agent through the navigator.modelContext API. One is you commanding your site; the other is other people's agents using it.
Can I edit my website with ChatGPT or Claude through MCP?
Yes, if your platform runs an MCP server. You connect your own account, authorize the link, and describe changes in plain language. How much you can actually change depends on the platform: some limit the assistant to a block editor, while an MCP-native CMS like Devani lets it edit both content and full-page layout in the visual builder. AI website builders covers how that fits the wider category.
Do I need to be a developer or use API keys?
Not for the editing side. Because the assistant is your existing ChatGPT or Claude, you authorize the connection once and work in conversation, no keys to paste and no code to write. Some developer-oriented servers do expect tokens and technical setup, which is worth checking before you assume it'll be one-click.
Is it safe to let an AI edit my live site?
The connection is authenticated and the server controls exactly which actions are allowed, so the assistant can't do more than the platform lets it. The bigger safety net is rollback: a setup that snapshots before every change means a bad edit is one click to undo. On a platform without that, be more cautious.
Which website platforms support MCP?
By mid-2026, WordPress.com and Pressable, Webflow, Wix, enterprise systems like Sitecore, Optimizely, and Umbraco, and Devani, among others. WordPress also has the Vibe AI plugin. They differ most in how deeply the assistant can operate the site once connected, so read past the "we have MCP" headline. For the WordPress picture specifically, the WordPress review covers where its AI editing works and where it stops.
Strip the hype and MCP for websites is a plumbing story: a standard connection between an assistant and a system that agreed to expose actions to it. That's it, and that's plenty. The useful question isn't whether a platform has MCP, since most now do, but whether the connection reaches all the way into how you build, and whether there's an undo button when the assistant guesses wrong. Answer those two and the marketing sorts itself out.
See what MCP-native actually feels like.
The fastest way to understand editing by conversation is to do it once. Connect your own ChatGPT or Claude and change a real page, rollback included.