How I build it · 01 / 04

List what the team does by hand.

A week of real site work: publishing posts, fixing typos, uploading images, checking SEO, copying leads into the CRM. Each task written down with how long it takes.

Design the tools and the limits.

Ten to twenty tools named after real jobs, like “create draft” or “find internal links”. Who can do what, what stays a draft and what needs a person to approve.

Build it on your stack.

A small server on your WordPress, Webflow, HubSpot or headless site, with an audit log, a note for every live change and undo from revisions. Connected to Claude, ChatGPT or both.

Hand it over with real tasks.

We run your team’s own jobs through it on the call. You get docs, example prompts and a log of everything it did.

Site jobs your team does by hand0 by handMapping
Step 1 · map0 / 5 automated
  • Publish a postMapping…
  • SEO checkMapping…
  • Internal linksMapping…
  • Live editMapping…
  • Undo a changeMapping…
  1. 01 / 04

    List what the team does by hand.

    A week of real site work: publishing posts, fixing typos, uploading images, checking SEO, copying leads into the CRM. Each task written down with how long it takes.

  2. 02 / 04

    Design the tools and the limits.

    Ten to twenty tools named after real jobs, like “create draft” or “find internal links”. Who can do what, what stays a draft and what needs a person to approve.

  3. 03 / 04

    Build it on your stack.

    A small server on your WordPress, Webflow, HubSpot or headless site, with an audit log, a note for every live change and undo from revisions. Connected to Claude, ChatGPT or both.

  4. 04 / 04

    Hand it over with real tasks.

    We run your team’s own jobs through it on the call. You get docs, example prompts and a log of everything it did.

  5. First week with the connector0 by handAutomated
    Step 4 · hand over5 / 5 automated
    • Publish a post✓ One request
    • SEO check✓ On every draft
    • Internal links✓ Suggested
    • Live edit✓ Reported
    • Undo a change✓ From revisions
    Test request: a draft written, SEO-checked and sent for review from one chat message.
Mapping: 0
What’s covered

Six parts of a safe connector.

Each one built and tested on your site, with what goes wrong when it’s missing.

PartWhat I buildWhat goes wrong without it
ContentDrafts, edits, media and publishing for pages, posts and CMS itemsAI writes markup the theme can’t render
PermissionsRoles, draft-only mode and approval for anything liveOne admin password shared with a chatbot
Audit logEvery call logged, every live change sent to Slack or TelegramNobody knows who changed the pricing page
UndoAny change restored from revisions with one requestA bad edit fixed by hand, from memory
SEO & linksTitles, descriptions, thin pages, internal links and content gapsNew pages live with no meta and no links
DataForm leads, GA4 and Search Console, read-onlyNumbers pasted into the chat by hand
Pricing

Fixed scope. Fixed price.

Before you sign, we agree the list of tools and what each one may change. You see every tool tested at handover.

Connect$1,5001 site · drafts only · 1–2 weeks
  • Up to 8 tools on one site
  • Drafts only, nothing goes live
  • Works with Claude and ChatGPT
Get a fixed quote
Site + CRMComplete$5,500Site, CRM and analytics · 4–5 weeks · one assistant for all of it
  • Everything in Fullcontent, media, live edits, undo, SEO checks
  • CRM toolsread form leads, update contacts in HubSpot
  • AnalyticsGA4 and Search Console, read-only
  • Automationsscheduled checks and a weekly report
  • Team rolesseparate access for editors, marketing and sales
Get a fixed quote
Prices exclude VAT · Claude or ChatGPT seats are billed by their providers

What an MCP server is

MCP, the Model Context Protocol, is an open standard for connecting AI assistants to other software. Anthropic published it in late 2024, and Claude, ChatGPT, Microsoft Copilot Studio, Cursor and VS Code now support it.

An MCP server for your website is a small program that gives the assistant a list of tools: create a draft, update a page, upload an image, check SEO. The assistant reads the list, picks the right tool for the request and calls it. It can’t do anything the tools don’t allow.

What your team can do with it

Typical requests sound like this: “Write a draft about our new pricing and link it from the pricing page.” “Find every service page without a meta description and fix them.” “Which questions did visitors ask in the chat this week that we don’t answer anywhere?” “Show me last week’s form leads by source.”

Each one used to mean logging in, copying text between tabs and remembering which plugin does what. Now it’s one message, and the result is a draft you review.

Why not just give AI your admin password

An admin login lets an assistant do anything, and you can’t see what it did. A good MCP server works the other way round: it does only what each tool allows, and leaves a trail.

Drafts are the default. Anything live goes through a tool you chose to allow, is logged, and is reported to Slack or Telegram with what changed. Every edit can be undone from revisions. Text written by site visitors, like chat messages or form entries, is treated as data, never as instructions, so nobody can steer the assistant through your contact form.

If your site is old

An outdated site is often the best case for this. The design may be dated, but the content still needs to change every week, and every change is slow because nobody remembers how the theme works.

The MCP server sits next to the site and talks to it through the platform’s API. Most WordPress sites have one built in. If a platform has none, I add a small one. Your team gets a modern way to edit an old site, and you can plan a rebuild when it suits you, not because editing hurts.

The one I built for my own site

This site runs on a headless WordPress with a Next.js front end, and I manage it through my own MCP server. It has 16 tools: drafts, live edits, publishing, media, revisions, SEO checks, internal links, content gaps and the questions visitors ask in the site chat.

Every live change sends me a Telegram note with what changed. Any change can be undone from revisions. After the assistant reads text from visitors, it can save drafts but can’t change the live site for an hour. This page was drafted through it.

After handover

You own the code. It lives in your repository, with tests for every tool and a short document that explains what each one does and who can use it.

AI tools and platforms change often. When one does, or when your team wants a new tool, I quote the change at a fixed price, like any other job.

Want to see it working?I’ll show you my own site’s server on a short call: a draft written, checked and reported live.Ask for a demo
MCP server FAQ

MCP, answered.

What MCP is, which AI tools it works with, security and platforms.

Something else?Write a line about your site. You’ll get a reply from me within one business day.

A small program that gives an AI assistant a set of named tools for your site, like “create draft” or “check SEO”. The assistant can only do what those tools allow, nothing else.

Anything that speaks MCP. Today that means Claude, ChatGPT on plans with custom connectors, Microsoft Copilot Studio, Cursor and VS Code. One server works with all of them.

Safer than a shared admin password. It writes drafts by default, live changes need a tool you choose to allow, every change is logged and reported, and any edit can be undone from revisions.

WordPress, Webflow, HubSpot CMS and headless sites on Next.js. For other CMSs I quote after a short look at their API.

Usually, yes. Most old WordPress sites already have an API, and if a platform has none I add a small one. It’s often a cheaper first step than a rebuild.

No. They type requests in plain words. You get short docs and example prompts for your own tasks, and we run them together on the handover call.

Admin access to the CMS, a staging copy if you have one, and a list of who will use the assistant and what each person should be able to change.

Often done together