Your AI can write you a website. Ask it, and you will get real HTML, real styling, the lot. Then you ask it to put the site online, and it says it cannot.
That gap is what MCP closes.
The short version
An MCP server is a set of abilities you hand to your AI assistant. Connect one, and your assistant can go and do the thing instead of describing it.
MCP stands for Model Context Protocol. It is an open standard — not a product, not something you buy — for describing tools to an AI in a way it understands. The AI companies agreed on one format so that every service does not have to invent its own.
You do not need to know any of that to use it. Practically, it works like adding an app to your phone: you add it once, and from then on your assistant can use it.
What was actually missing
Here is the thing most people get wrong, and it is not their fault.
When your AI "cannot publish your website", it is not being lazy and it is not too stupid. Reading a page and sending something to a website are two different abilities, and a chat in a browser tab usually only has the first one.
It is a bit like the difference between someone who can read a recipe out to you and someone who is standing in your kitchen. Same knowledge. Very different amount of help.
An assistant with no MCP connection can:
- read instructions on a website
- write you a complete, working site
- explain exactly what you would need to do next
It cannot press the button. You have to.
What changes when you connect one
Once your assistant has an MCP connection, it stops handing you homework.
The difference shows up most clearly when something goes wrong. Without a connection, a broken website looks like this: you publish it, you visit it, it is blank, you copy the error, you paste it back into the chat, it suggests a fix, you apply the fix, you publish again. Four round trips and you are the one carrying messages between two things that cannot talk to each other.
With a connection, the assistant reads the error itself and fixes it. You asked for a website; you get a website.
That is the whole point. Not that it is faster to type — that it removes you from the middle of a conversation you were never needed in.
Which AI assistants support MCP
Support falls into two groups, and the difference matters more than the brand name.
Assistants that run on your computer. These have supported outside tools the longest, because they were already able to touch things — your files, your terminal. Google Antigravity, Cursor, Claude Code, OpenCode, Gemini CLI, Codex and Devin are all in this group. If you already use one of these, you are sorted; Antigravity is free, which makes it the easiest place to start if you are not paying for an AI tool today.
Chat apps with connectors. This is the newer group and the more interesting one, because nothing gets installed. Claude and ChatGPT both let you add an outside service from their settings, on their paid plans. You add an address once, and the assistant can publish from inside the conversation.
Plain chat in a browser tab, with nothing added. No support. It will write you a beautiful website and it cannot put it anywhere. This is the case that confuses everybody, because it is the same brand and often the same web page as the version that works.
We keep a page comparing the specific products, because the siblings genuinely get confused — the Gemini chat and Gemini CLI are different things, and so are Claude in a browser and Claude Code. If you are not sure which you have, start there.
How to connect DashPloy
DashPloy runs an MCP server so your assistant can publish websites for you. The address is:
https://dashploy.com/mcp
Three steps, and none of them involve a terminal.
1. Find where your app lists connectors. It might be called connectors, MCP servers, integrations or tools, depending on the app. Add a new one and paste the address above.
2. Choose to sign in. If it asks how it should connect, pick the option that signs you in rather than the one asking for a key, a password or a token. There is nothing to paste.
3. Press Connect. Your app sends you to a DashPloy page that lists exactly what it is asking to do — create websites, publish them, connect domains. Read it, press Connect, and you are done.
After that, ask for a website the normal way. "Build me a one-page site for my bakery and publish it" is a complete instruction.
You can withdraw the connection whenever you like from Connected apps, and doing so cuts off its access immediately.
If your app asks for a key instead
Some apps have no sign-in option. Those want a header, and the name to use is
X-API-Key, with your code from your dashboard as the value.
Use that name even where the app offers you Authorization as well. Some apps
keep Authorization for their own sign-in and quietly discard whatever you put
there — and the connection still looks perfectly fine afterwards. It says it is
connected, it lists everything it can do, and it fails at the first attempt to
publish. We learned that the hard way rather than from any documentation.
Check it actually worked
Ask your assistant: "call whoami on DashPloy".
If it comes back with your email address and your plan, the connection is real.
If it says no token was sent, the code is not reaching us — go back and use
X-API-Key.
This is worth doing at setup rather than later. A connection that looks fine but is not will otherwise reveal itself at the worst moment, which is halfway through building something.
Is it safe?
Reasonable question, and the honest answer has two halves.
What it can do: publish websites to your account, change them, read their logs, connect domains you own, and delete websites. That is real access, and the page you approve says so plainly rather than burying it.
What it cannot do: sign in as you anywhere else, see your password, change your plan, or touch your billing details. And it keeps nothing after you disconnect it.
The sign-in route is safer than pasting a code, for a mundane reason: nothing gets copied. A code pasted into a chat sits in that conversation's history, and in the next one, and in the screenshot you took. A connection you approved has none of that, and you can end it from one page.
Do you actually need this?
Honestly, not everybody does.
If you publish a website once and change it twice a year, pasting a line into your assistant works fine and always will. Nothing about it is second-class — it is the same platform underneath.
Connect an MCP server if you are building more than once, if you want your assistant to notice and fix its own mistakes without you relaying error messages, or if you have never opened a terminal and would rather it stayed that way. That last one is the case we built this for. It is the shortest path that exists from "I need a website" to a working address, and it does not require you to install a single thing.
If you want the technical side — the transport, the tool list, what to do when a connection fails — that is in connecting DashPloy with MCP and in the documentation.