Build a Website With Claude, Codex or Cursor: MCP Guide

You can build and run a real website from the AI you already use. Not a code export, not a chat panel bolted onto a builder, but your own Claude, ChatGPT, Codex or Cursor opening your site, reading every page, writing a section and publishing when you say so. The thing that makes this possible is MCP, and this guide explains how to work that way, from the first connection to the publish call.

It is written for people who run a site, or build sites for others, and who already pay for an AI assistant. Everything in it is how we run beste.co itself. The six posts linked from it go deeper into each part.

What MCP is, in one paragraph

MCP, the Model Context Protocol, is an open standard for connecting AI applications to outside systems. A product runs an MCP server that describes what it can do as a list of tools, each with a name, a description and the inputs it accepts. The assistant reads that list and calls the tools when a task needs them. The protocol documentation explains the client and server roles in detail. Claude, ChatGPT, Claude Code, Codex, Cursor and many others can act as the client.

For a website builder, this changes one thing that matters: the AI stops writing instructions for you to follow and starts doing the work inside the product, on the real site.

What changes when your builder has an MCP server

Most AI website builders give you a chat box inside their product. You type, the builder's own AI changes the site, and the builder bills you for it in credits or tokens. That works, and it has three limits.

First, you are paying for AI twice: once for the assistant you already use, and again for the builder's. Second, the builder's AI only knows the builder. It cannot read your brand document, your old site, your product spreadsheet or the notes you keep in your own tools. Third, the conversation lives inside the builder, so it is one more place where decisions get made and lost.

With an MCP server, the roles flip. Your assistant stays where it is, with everything it already knows and every other tool it is connected to, and the builder becomes one more set of tools it can call. The context is everything you have given your assistant. The result is a normal site in a normal account, not a copy or an export. The full comparison is in MCP Server vs Built-In AI Chat.

Not every MCP server works the same way

This is the distinction most comparisons miss. Builders now publish MCP servers in two different designs.

Servers that forward to the builder's own AI. Your assistant calls a tool such as "send a message to the project", and the builder's agent makes the change, exactly as if you had typed into its chat box. Lovable's server works this way: its send_message tool sends a chat message to the project's AI agent, and its documentation says standard Lovable credits apply to create_project and send_message calls. You get to stay in your own client, but the work, and the bill for it, is still the builder's AI.

Servers that expose the editing actions themselves. Your assistant reads the page, picks the section, writes the copy and saves it, one tool call at a time. There is no second AI in the middle. Beste's server is built this way: tools like get_page_markdown, update_section, create_post and publish are the same actions you take in the editor.

Neither design is wrong. They answer different questions. The first lets you drive another product's AI from your chat. The second lets your own AI do the work, with your context and your rules, and leaves the builder in charge of keeping the result safe. When you compare builders, ask which one you are getting.

What you can do over MCP

On a builder with a complete, tool-level MCP server, the AI can do what you can do in the editor. For Beste that means:

Build a site from a prompt. Pick sections, write copy, set up pages, navigation and footer.
Edit what is there. Read a section as written, change one sentence, add an item to a list, swap an image.
Change the design when asked. Theme, fonts, colors, site look and animations, as a separate request from content.
Run the blog. Create, write, update and list posts, set authors and tags. See Running a Blog Through MCP.
Keep the site current. Prices, hours, team members and products, as the business changes. See Keeping a Website Current With an AI Agent.
Add languages. Add a language, pull every string, write the translation back, set local URLs. See How We Built a Multilingual Website With MCP.
Handle SEO and redirects. Titles and descriptions per page, robots.txt, a redirect when a URL changes.
Manage products and forms. Catalog items, prices and variants, and the forms that collect leads.
Track the clicks that matter. GA4 or Meta pixel events on specific links.
Preview and publish. Open the draft as a visitor would, then publish when you ask.

The Beste server has more than 70 tools covering all of this, each labeled read or write, listed on the MCP server page.

Choosing a client

The same server works from several places, and the right one depends on where you already work:

Claude on the web, desktop or phone through the Connector for Claude. Nothing to install. Good for content work, quick edits and anything you want to do from a phone.
ChatGPT through the Plugin for ChatGPT. Build a new site or edit the one you run, with the result in a live preview next to the chat.
Claude Code or Codex in a terminal. One command, npx beste-agent@latest, signs you in. Good for long sessions, bulk work and anything you want to repeat.
Cursor, Gemini CLI or Grok, connected from the Beste dashboard, for people who already live in those tools.

How to pick, and how each one behaves in practice, is in Claude Connector, Claude Code or Codex: Which One to Build Your Website With.

How a session goes

Here is the shape of a working session, the way we run them on our own site.

1. Connect, once

In Claude, add the connector and approve the projects it may work on. In ChatGPT, install the plugin and sign in. In Claude Code or Codex, run npx beste-agent@latest, which signs you in through the browser and sets up the server. Access is per project and you can remove it at any time. After that, the tools are simply there in every session.

2. Let it read before it writes

Ask the assistant to list your sites and read the page you want to change. A good MCP server returns the page as text, section by section, with ids. This step is what makes the edit land in the right place. An assistant that skips it guesses, and guesses badly. Reads should be free, so there is never a reason to skip them. On Beste they do not count toward any limit.

3. Give it the rules

Tone, glossary, what it must never change, how it should handle facts it does not know. Put these in a document the assistant reads at the start of every session: a CLAUDE.md or AGENTS.md file for Claude Code and Codex, or the project instructions in Claude and ChatGPT. Page 30 will sound like page 1 only if the rules were written down. What goes in that document is the subject of Writing Rules for Your Website Agent.

4. Work in the draft

Ask for the change. The assistant calls the tools, the change lands in the draft, and nothing reaches visitors. Ask for several changes in a row, then read the draft once. Batching is faster for you and cheaper on your attention.

5. Review as a visitor

Ask for the preview and open it. Read the page the way a visitor would, on a phone if that is where your visitors are. Fix what reads wrong by asking for the specific fix ("shorten the second heading"), not by explaining what went wrong in the abstract.

6. Publish, in your own words

Say publish. The assistant should publish only when you ask, and only the draft you just read.

That loop, read, rules, draft, review, publish, is the whole method. Everything else is detail.

What it looks like in practice

A real request from our own week, and the calls behind it:

"The Pro price on the pricing page should say billed annually under it. Check every page that mentions the Pro price."
The assistant calls list_pages, then get_page_markdown on each page that could mention pricing, and finds the price in three places: the pricing table, an FAQ answer and a comparison section.
It reads each section with get_section, changes the line with update_section, and reports the three changes.
We ask for the preview, read the three spots on a phone, and say publish.

Four messages from us, about two minutes, and the part that used to go wrong, the third place nobody remembered, did not.

Where the AI is better than you

Volume and consistency. We rebuilt beste.co in three languages this way, 46 pages in each, and the assistant never skipped an alt text, never left a button untranslated and never forgot the FAQ at the bottom of a page. Once a word choice was set, it kept it across the whole site. A person doing this by hand gets tired by page 15. The assistant does not.

It is also better at the tedious technical middle: setting a localized slug on every page, writing a meta description for every page, adding a redirect for every old URL. These are tasks where the risk is forgetting one, and forgetting is the one mistake it does not make.

Where you still decide

Three kinds of fixes come up in every session, and they are the reason a person stays in the loop:

Facts. Prices, hours, names, claims. The assistant drafts them and you check them against the real material. It will write "open until 8pm" with total confidence.
Voice. A headline can be correct and flat. Short marketing lines need an ear, and the ear is yours.
Scope. The assistant will happily add a page you did not ask for because it seemed useful. Decide the scope before you start, and say so in the rules.

The pattern is simple: the AI drafts and suggests, you decide, and you read every page once before it goes live. Reviewing is a much smaller job than building.

What keeps it safe

An assistant with write access to your live site needs guardrails, and they should live in the server, not in your hopes. Instructions in a rules file shape what the assistant tries. Anthropic's own documentation describes CLAUDE.md as context the model reads, not configuration it is forced to follow, which is exactly why the hard limits belong in the server. These are the ones we built into Beste's, and the ones to look for in any builder's:

Draft first. Every change goes to the draft. Visitors see nothing until you publish.
Publish only on request. The publish tool runs after your request, in your own words, not as a side effect of another task.
Tools that say what they change. Each tool is labeled read or write, so the assistant can tell a safe call from a destructive one before it makes it.
Errors it can fix. A rejected request returns the field path and the expected values, so the assistant corrects itself instead of asking you to read an error message.
Batch undo. Changes land in batches, and most batches can be undone with one call.
A second step for deletion. Deleting a blog post needs a confirming call. Undo covers most edits, but a deletion may not be fully restorable, so the server makes the assistant stop and ask.
Access per project. The assistant works only on the projects you approved, and you can remove access at any time.
A ceiling on runaway loops. Requests are rate limited per hour, so a confused session cannot hammer the site.

On your side of the connection, the usual MCP advice applies. Anthropic's guidance on custom connectors is to connect only to servers you trust, review the permissions a server asks for, and keep an eye on what the assistant does with them. Prompt injection is real: a page or file the assistant reads can contain text that tries to steer it. A server that keeps everything in a draft until you publish limits what any one bad instruction can do.

What it costs

On the AI side, nothing beyond the plan you already pay for. Claude's custom connectors over remote MCP are available on its Free, Pro, Max, Team and Enterprise plans, with Free limited to one custom connector, per Anthropic's support article. Codex is included in the ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise plans listed on OpenAI's Codex pricing page, with usage limits that depend on the plan and the size of each task.

On the Beste side, reads never count. A Free project gets 150 AI agent edits a month, and Pro, Team and Business include unlimited agent edits. Beste's own AI credits pay for image generation only. Pro is $15 a month billed annually, or $20 billed monthly. Full details are on the pricing page.

Mistakes we made in the first month

So you can skip them:

Starting without rules. The first ten pages were fine. Page eleven used a word we had never used. The rules document started that afternoon.
Asking for too much in one request. "Rebuild the features page and translate it" produced a long session that was hard to review. One change per request, then a batch review, works better.
Reviewing in the chat instead of the preview. A paragraph that reads well in a chat window can wrap into a wall on a phone. Always read the preview.
Letting it guess facts. It filled a placeholder with a plausible number once. Now the rule is that unknown facts stay as visible placeholders.

How this works in Beste

Beste runs an MCP server with more than 70 tools for sites, pages, sections, shared navbars and footers, blog posts, authors, products, forms, images, redirects, languages, design and publishing. The same tools serve the Connector for Claude, the Plugin for ChatGPT and the MCP server for Claude Code, Codex, Cursor, Gemini CLI and Grok. Bestie, the assistant inside the editor, uses them too, so a site started in one place can be continued in another.

The site you are reading was built, translated and is kept current through these tools, with the decisions made by people and the volume done by the assistant.

If you want to see why this matters for AI search as well as for building, our Agent-Ready Websites page covers the other direction: how assistants read your site once it is live.

Every post in this guide

Sources

Checked in October 2026. Plans and limits change, so check each page before you decide.

Building and Running a Website From Claude, Codex or Cursor: The MCP Guide