How I Build Websites Fast With Claude and MCP

I put off writing this one for a while because most "I build sites with AI now" posts are either breathless ("I made a startup in a weekend!") or dismissive ("it's just autocomplete"). The truth is duller and more useful than both. Over the last year Claude has quietly become the tool I reach for most, and MCP is the part that turned it from a clever chat window into something wired into how I actually work. So here's the honest version — what it does for me, what it doesn't, and where I still won't let it drive.

MCP is the part that made it click

For a long time using AI meant copy-pasting: grab an error from my terminal, paste it into a chat, paste the answer back, repeat. Fine, but it's a lot of tab-switching, and the model never really knew what my project looked like.

MCP — the Model Context Protocol — is what closed that gap for me. It's basically a standard way to hand Claude real tools and real context instead of me narrating everything by hand. In practice that means Claude can read the files I'm working in, run the commands I'd run, and reach into the services I've connected. It stops guessing what my codebase looks like and just… looks. That single change is most of the speed-up. Not smarter answers — fewer round trips.

A normal build, start to finish

Here's roughly how a small client site goes now. It's not magic, it's just fewer dead minutes.

• I talk through the structure first — pages, sections, the content model — in plain language, and get pushback before a single component exists.
• Once the shape is agreed, I have it scaffold the repetitive parts: layout shells, a CMS structure, the twelfth card component that's basically the eleventh one.
• I stay in the driver's seat for anything with taste in it — spacing, motion, the actual feel of the thing.
• At the end I point it at the whole diff and ask what I got wrong, which is where it earns its keep.

The pattern underneath all of that: I use it hardest at the two ends. The boilerplate start, and the review at the finish. The middle — the part that's actually design — is still mostly me.

Where it genuinely saves me hours

The wins are unglamorous, which is exactly why they add up. Boilerplate I've written a hundred times — form handling, responsive grid scaffolding, a config file I always have to look up — comes out in seconds instead of fifteen distracted minutes. It's a good rubber duck: explaining a bug out loud to something that talks back often surfaces the answer before it does. And it's fast at the tedious refactors, like renaming a prop across forty files, that I'd otherwise keep putting off.

On one recent build I timed it, roughly. The parts I'd normally lose an afternoon to — wiring up a contact form, setting up the blog structure this very post lives in, chasing a layout bug across breakpoints — took about a third as long. Not because the code was cleverer. Because I never left the editor and never lost the thread.

Where I still don't let it drive

This is the part the hype skips. AI is confidently wrong often enough that trusting it blindly is how you ship something broken with great commit messages. So there are places it doesn't get the keys.

• Design taste. It'll give you the safe, average version of a layout every time. Average is the one thing a client isn't paying me for.
• Anything security-shaped — auth, form handling, anything touching a key. I read every line of that myself, twice.
• Copy that's supposed to sound like a real person. It drifts toward the same tidy, listy, faintly hollow voice, and people can feel it even when they can't name it.
• The final call on whether something is actually good. That's the whole job, and it's not delegable.

That third one is worth sitting with, because it's the reason this site's writing is written by me and not generated. AI copy has a texture — everything balanced, every paragraph the same length, a little too eager to wrap up with a neat bow. It reads fine and lands nowhere. If a page is meant to sound like a person, a person has to write it. The AI can catch my typos.

So is it actually faster?

Yes — but not in the way the demos promise. It doesn't build the site for me. It removes the friction between the parts I care about, so I spend my hours on the decisions that need a human and almost none on the setup that never did. The work that's left is more concentrated, and honestly more fun. The tool's fast. Knowing which parts to hand it, and which to keep — that's still the job.

Want a site built this way? Let's talk