What is Claude Code?
Claude Code is an agentic coding tool from Anthropic - an AI that works inside a real codebase the way an engineer does, rather than suggesting snippets in a chat window. This is the plain-English version: what it is, why everyone is talking about it, and what it should mean when an agency says it builds with it.
In short: Claude Code is an AI coding agent made by Anthropic, the company behind the Claude models. You give it a goal in ordinary language; it reads the project, writes and edits the code, runs the tests and ships the change. It isn’t a website builder and it isn’t a chatbot. For a business buyer the significance is simple: the labour of building and changing a website has collapsed in cost and time - and an agency that uses the tool properly can pass that on.
The plain-English definition.
Claude Code is made by Anthropic, the AI company behind the Claude family of models. It started life in a developer’s terminal - the black window of typed commands - and now also runs inside code editors, on the desktop and on the web. Wherever it runs, the job is the same: it operates on a real software project, end to end.
The easiest way to place it is against the things it isn’t. ChatGPT-style assistants answer questions; website builders like Wix hand you templates to drag around. Claude Code is neither. It’s closer to hiring an engineer: you describe an outcome - “add a pricing page”, “make the forms tell us which service was picked” - and it does the engineering, reading the existing code, making the changes and checking they work.
That’s why it matters well beyond the software industry. Websites are code. When the cost of writing and changing code drops by an order of magnitude, the cost of a good website should drop with it. Whether it actually does depends on who’s holding the tool - which is where the rest of this page, and our approach to building on it, come in.
What “agentic”
actually means.
The word behind the hype, and why the review step is the point rather than a limitation.
“Agentic” is the industry’s word for a tool that carries out a task rather than assisting with one. The first wave of AI coding tools were assistants: autocomplete on steroids, suggesting the next line while a developer typed. Useful, but the developer still did the work.
An agent works differently. Given a goal, Claude Code will:
- read the codebase to understand how the project fits together
- plan the change and set out what it intends to do
- write and edit the files itself
- run the code, the tests and the build to check nothing broke
- present the finished change for a human to review and approve
It loops - build, check, fix - until the job is done or a person steps in. That human review step is the point, not a limitation: the machine does the labour, a person confirms the result is right. On our builds that person is the same senior lead who did the research and directed the design, which is what keeps AI speed from turning into AI slop.
Where Claude Code came from.
Anthropic released Claude Code in February 2025 as a research preview, alongside a new generation of Claude models - the first strong enough to carry real engineering work end to end. By May 2025 it was generally available, and through 2025 and 2026 it spread from the terminal into code editors, the desktop and the browser.
The reason it became famous is uncomplicated: it was the first mainstream tool where the AI actually did the work. Developers stopped copying suggestions out of a chat window and started delegating whole tasks. By 2026, agentic tools are how a large share of professional software gets built - and websites are software. Anthropic runs a partner programme around Claude for exactly this kind of commercial use; Web Hero is an Anthropic Claude Partner, and it’s the platform every one of our code builds ships on.
What it changes if you’re
buying a website.
None of this matters to you as a tool. It matters as economics. Four things change when a site is built this way, each for a mechanical reason rather than a marketing one.
Speed: live in days
The build stops being the bottleneck. Research, design and sign-off still take human time; the engineering that used to take a quarter compresses into days.
The cost of change collapses
The old model billed every change at £350–£500 because a person rebuilt it step by step. With an agent inside the codebase, a change is described, made and shipped - usually the same day.
No platform tax
The output is plain code in a repository, not pages locked inside someone’s builder. No monthly platform fee, no export wall: if you leave, you leave with everything.
A higher quality ceiling
Google Lighthouse flags everything slow or inaccessible about a site. On a template, much of it can’t be reached. In plain code every flag is fixable, which is why our builds score 100 across all four categories.
We’ve broken the numbers down properly - what the tool costs to subscribe to, and what a professionally built site costs - on the Claude Code pricing UK page.
What it doesn’t change.
You can outsource the thinking. You can never outsource the understanding. Claude Code writes code; it does not know your buyers, your margins or which enquiry is worth the most to you. Strategy, positioning, design judgement and sign-off stay human - and a tool this fast in careless hands simply ships careless work faster.
So judge any “AI-built” website the way you’d judge any website: on the thinking behind it, the results it gets and the people accountable for it. And if what your team actually wants is hands-on, drag-and-drop editing, an agentic build isn’t the answer at all - that’s what Duda is for, and we’ll say so on the call.
Claude Code vs the tools
you’ve heard of.
The names get used interchangeably in coverage. They shouldn’t be - they do different jobs.
ChatGPT
A general-purpose assistant. It can draft code in a chat window, but it doesn’t live inside your project or ship changes. OpenAI’s real counterpart to Claude Code is Codex, its own coding agent.
GitHub Copilot
The best-known coding assistant: it suggests code while a developer types inside their editor. An accelerator for hands-on programmers, not an agent you can delegate a task to.
Cursor
An AI-first code editor. Powerful for engineers who want AI woven through hands-on-keyboard work - a different job from an agent that carries the task out itself.
We compare all four properly, including where each one genuinely wins, in Claude Code vs Cursor, Codex and Copilot.
What “built on Claude Code”
should mean.
“AI-built” is already appearing on agency websites, and it can mean anything from everything described above to a template with a chatbot bolted on.
Four questions cut through it:
- Who owns the repository afterwards - you, or the agency?
- What do the builds actually score on Google Lighthouse?
- What does a change cost after launch, and how long does it take?
- Who does the thinking - is a senior person accountable for strategy and design, or is the AI unsupervised?
Our answers are set out in full on Built on Claude Code: the repo sits in your name, every build targets Lighthouse 100s, changes ship by asking - usually the same day - and one senior lead directs research, design and build. This very site is built the same way, and the same foundation carries our wider AI agency work, including getting clients found in AI search as well as Google.
The proof is shipped work, not claims: Cademi, a Claude Code rebuild that reported 3.4x demo requests from the engagement; Avago, an end-to-end build spanning site, API, portal and brand; and LoadSnap, a UK waste-compliance product with a DEFRA Digital Waste Tracking integration. And if you already have a website, it doesn’t start from a blank page - we migrate sites from WordPress, Wix, Squarespace and the rest without losing rankings.
Fair questions.
Claude Code is an agentic coding tool made by Anthropic, the company behind the Claude AI models. It works inside a real software project: you describe a goal in plain language and it reads the code, makes the changes, runs the tests and presents the result for review. It runs in a developer’s terminal, in code editors, on the desktop and on the web.