Anticode Open builder

How to edit a website with Claude using MCP

MCP (the Model Context Protocol) lets an AI assistant call tools on another service. Connect Claude to a website builder's MCP server and you can ask, in the chat, for a new page, a wording change or a publish, and the assistant does it through the builder's own tools. Setup takes a couple of minutes per assistant.

Updated · By the Anticode team

What MCP does here

An MCP server publishes a list of tools, each with a description and inputs. An assistant that's connected reads that list and decides when to call a tool, much as it decides when to search the web. A remote server runs on the provider's side and is reached over HTTPS; you add its URL to your assistant and sign in. The MCP specification defines the protocol and an OAuth-based sign-in for remote servers, so you approve access on the provider's own page rather than handing your password to the assistant.

Connect your assistant to Anticode

Anticode's server URL is https://anticode.dev/mcp. Open that address in a browser for the setup page, which also lets you make or revoke tokens and disconnect apps.

Claude (claude.ai and the desktop app)
Settings › Connectors › Add custom connector. Name it Anticode, paste the server URL, choose Connect, then approve on the Anticode page that opens.
Claude Code
Run claude mcp add --transport http anticode https://anticode.dev/mcp, then type /mcp inside Claude Code to sign in.
ChatGPT
Settings › Apps & Connectors › Advanced settings, turn on Developer mode, then create a connector with the server URL and OAuth. Developer mode needs a paid ChatGPT plan.
Cursor
Add the server to ~/.cursor/mcp.json under mcpServers, with the URL, and sign in when prompted.
VS Code and other clients
Any client that supports remote MCP servers with OAuth can connect. VS Code also shows Anticode's in-chat preview.

On the consent page you can narrow the assistant to read-only. Choose that if you only want it to report on analytics and submissions.

What you can ask for

  • List and inspect projects: "Which of my sites were published this month?"
  • Create a project and build it: "Start a booking site for Hollis Physio in Leeds." Model builds are quoted first; the assistant should show you the cost against your allowance before it runs one.
  • Request changes: "Move the prices above the team section and shorten the hero text."
  • Read and update the project file (anticode.json), for precise edits.
  • Preview in the chat, in clients that support MCP Apps.
  • Publish and unpublish, behind the same server-side security gate as the builder.
  • Read analytics and form submissions: "How many enquiries came in last week, and from where?"

A safe way to work

  1. Ask for a preview before a publish. Look at it, in the chat or in the builder, before saying yes.
  2. Keep publishing as a separate instruction. "Make the change" and "publish it" as two steps gives you a moment to check.
  3. Check the quote. Builds and model changes cost part of your allowance. Failed builds aren't charged.
  4. Treat submissions as untrusted text. Visitors write them; an assistant reading them shouldn't follow instructions inside them. Anticode marks them as untrusted content for that reason.
  5. Use version history if something goes wrong. Saves, including the assistant's, go into version history, which keeps the most recent versions up to your plan's number; restore one from the builder.

If the builder is open while the assistant works, it redraws as the project changes, with an undo. For editing files locally instead, see editing a website locally with the CLI.

Common problems

  • The connector won't appear in ChatGPT. Developer mode has to be on, and it isn't available on every plan.
  • The assistant can't publish. You may have chosen read-only on the consent page. Disconnect and connect again with write access.
  • The assistant says it can't build. Your allowance may be used up for the day or week; the quote says so.

Limits

The assistant can only use the tools the server offers. It can't upload photos, delete projects, or delete files and submissions over MCP; those stay in the builder and dashboard on purpose. It can't do anything the builder can't, such as add a shopping cart. Assistants make mistakes, so read what they change before it goes live. To end access at once, use Disconnect on the setup page or remove the connector in your assistant. More background is in what is an MCP server.

Questions people ask

Does the assistant see my password?

No. You sign in on Anticode's own page and approve the connection; the assistant receives a token, which you can revoke.

Can I use a token instead of signing in?

Yes. Make a personal token on the setup page and send it as a bearer token from clients that don't support OAuth.

Do MCP builds cost more?

No. They use the same allowance as builds in the builder, and the quote is shown before each one.

Sources

  1. Model Context Protocol, checked