SiteClaw: Why I Built an AI That Builds Websites

SiteClaw started because I was tired.

Not burned-out tired. Process tired. I've been building websites for 22 years. I've shipped them for architecture firms in Southern California, tourism companies in Hawaii, media empires doing $435K a month in ad revenue, and a modeling platform that became a TV show on the CW. I know exactly how to build a great website.

I also know exactly how painful the process is. The intake calls. The design rounds. The revision emails. The "can you make the logo bigger" conversations. The gap between what a client imagines and what they can articulate. Multiply that by 171 projects and you start fantasizing about a world where you could just tell a machine what to build and watch it happen.

So I built that world. Or at least I'm trying to.

What SiteClaw actually is

SiteClaw is a platform where you describe what you want, and an AI agent builds it in a live preview while you watch. Not a template picker. Not a drag-and-drop builder with AI copy suggestions. An actual agent that reads your brief, makes design decisions, writes production code, and renders the result in a live preview URL — then iterates based on your feedback in natural language.

You can import a GitHub repo and have the agent work on an existing project. You can connect your Vercel account and deploy with a click. The agent has access to a template system, so it starts from smart defaults instead of a blank canvas. It can run Notion imports, manage deployments, and — eventually — build and schedule marketing automations through Zapier.

That's the vision. The reality is messier.

The bugs that build character

Here is a partial list of things that have gone wrong:

The agent tells me it fixed the build error. It describes the fix in detail. Nothing has changed. The response came back in under a second — it didn't even look at the code.

I send the agent a screenshot of a white screen with 404 errors on CSS and JS files. The agent acknowledges the screenshot, then suggests a fix that has nothing to do with what's in the image. It's guessing. Confidently.

The live preview shows a bad gateway error because the agent process timed out during a long operation. I have to log in again, but by then all context is lost and no memory has been saved. I've fixed this bug eight times. It keeps coming back.

CSS and JS assets 404 after rebuilds. The preview URL works on the agent's end but not on mine. The agent claims the preview is live when it demonstrably is not — even though I explicitly told it, through its own system memory that it created itself, to verify builds before sending confirmations.

This is the reality of building with agents. The models are brilliant at reasoning and terrible at self-verification. They will hallucinate success.

What I learned the hard way

The breakthrough wasn't making the agent smarter. It was making the system around the agent more honest.

I ported from Timescale to Neon when costs got out of control. I built a hot-loading context system so agents can recover from crashes without losing the conversation. I'm implementing per-project vector memory so the agent can access recent messages on reconnection and automatically restore the live preview for the right project.

The template system means the agent doesn't start from nothing — it picks the right scaffold for the project type and builds from there. GitHub import lets you bring existing repos into the platform. The deployment pipeline goes from git push through Vercel deploy through Lighthouse audit, all automated.

None of this is glamorous. None of it demos well. But it's the infrastructure that makes the difference between a prototype that works in a screen recording and a product that works when real people use it.

Why I keep going

Because when it works — when you type "make the hero section taller and swap the CTA to a dark button" and watch the preview update in real time — it feels like the future arrived early.

I ran an agency for five years. I was the number one WP Engine partner two years running. I have built more websites than I can count. And I can tell you with certainty: the version of web development where you describe what you want and watch an agent build it is better than every version that came before.

It's not ready yet. The agent still lies sometimes. The preview still breaks. The context still gets lost. But every week it gets a little more reliable, and the gap between what it is and what it needs to be gets a little smaller.

I'm building in public. The bugs are real. The progress is real. Come watch.