What Is Claude Code? A UK Business Guide (2026)

Claude Code is Anthropic's command-line AI agent for software engineering. It reads entire codebases, writes and edits code across hundreds of files, runs tests, and ships features — with a human reviewing every change before it lands.

Last updated: April 2026 · Covers Opus 4.6, Sonnet 4.6, Haiku 4.5

What is Claude Code in plain terms?

Claude Code is a command-line tool. You install it, point it at a codebase, and give it tasks in plain English. It reads every file it needs, plans the change, writes the code, runs the tests, shows you the diff, and waits for approval.

It is not autocomplete. It is not a chatbot bolted onto an IDE. It is an AI coding assistant that operates the way a senior developer operates — except it works through your entire backlog without getting tired, distracted, or expensive. Where traditional AI code generation tools produce snippets, Claude Code produces finished, tested features.

Cursor and Copilot help a developer type faster. Claude Code does the work.

Here is what that looks like in practice. A developer opens a terminal, starts Claude Code, and types a task:

` $ claude

Add rate limiting to the /api/leads endpoint. Max 60 requests per

minute per API key. Return 429 with a Retry-After header when exceeded. Write tests.

Claude Code is working...

✓ Read src/app/api/leads/route.ts (existing endpoint) ✓ Read src/lib/rate-limit.ts (found existing limiter utility) ✓ Edited src/app/api/leads/route.ts — added rateLimiter middleware ✓ Created src/app/api/leads/__tests__/rate-limit.test.ts (6 tests) ✓ Ran npm test — 6/6 passing

Changed 2 files, created 1 file. Ready for review.

/diff

`

The developer reviews the diff, approves it, and the feature ships. Time from task to tested PR: twelve minutes.

What makes Claude Code different from other AI coding tools?

Claude Code is the only AI coding tool that operates at the whole-codebase level, plans multi-step work autonomously, and maintains persistent memory across sessions. It is the official CLI from Anthropic for agentic software engineering, designed by Boris Cherny and the Claude Code team. You run it in a terminal. It connects to the Claude model family — currently Opus 4.6, Sonnet 4.6, and Haiku 4.5, with Opus 4.7 on the near horizon — and performs multi-step work against your code.

"The terminal is not a constraint — it is the point. The terminal is where Git, deployment tools, and production infrastructure already live."

— Boris Cherny's design thesis for Claude Code

An IDE plugin can suggest code. A terminal agent can ship it.

Three things make it different from previous generations of AI coding tools:

It operates at the codebase level, not the line level. Most IDE plugins see the file you are currently editing, plus maybe a few adjacent ones. Claude Code reads the whole repository. It understands project structure, knows where the tests live, notices your conventions, and references your existing patterns when it writes new code.

It is agentic, not suggestive. You do not accept suggestions one line at a time. You give it a task. "Refactor the auth middleware to use the new token service." "Add rate-limiting to the public API routes and write tests." "Investigate why the Stripe webhook is dropping events and propose a fix." Claude Code plans the work, executes the steps, and returns with either a finished change or a report on what it found.

It has memory and tooling. Through a project-level CLAUDE.md file, it picks up your codebase conventions, banned commands, review standards, and preferred tools. Through the Model Context Protocol (MCP), it can connect to your databases, monitoring systems, internal APIs, and third-party services — all under your control, all with audit logs. In practice, this means your agent can pull live data from Salesforce, trigger a deployment pipeline, or query your production database without you copying between windows.

The result: you stop writing code line by line and start supervising an agent that ships whole features.

How does Claude Code compare to Cursor and GitHub Copilot?

Copilot helps you type faster. Cursor helps you edit faster. Claude Code finishes the ticket. The three tools sit at different points on the same curve.

Most serious UK teams end up using more than one of these. Copilot or Cursor for in-flow editing, Claude Code for the tasks that would otherwise take half a day. They compound rather than compete. A full head-to-head covering token costs, privacy posture, and task-routing guidance follows in the next piece in this series.

Does Claude Code actually work in production?

Yes — and the proof is public, named, and measured.

TELUS rolled Claude Code out to 57,000 employees across its Canadian operations. Pull-request turnaround dropped by 30%. The company has since expanded its Anthropic partnership to cover wider enterprise use.

IG Group — a FTSE 250 financial services business headquartered in London — reported that their analytics teams save 70 hours per week through Claude-powered automation. That is the equivalent of nearly two full-time employees per team returned to higher-value work.

Zapier deployed Claude internally at scale — more than 800 Claude-driven agents in production, automating work across engineering, marketing, and customer success. The volume of internal tasks completed via Claude has grown tenfold year-on-year.

Cox Automotive integrated Claude across VinSolutions CRM, Autotrader PSX, and Dealer.com. Consumer lead responses and test-drive appointments both doubled. AI-generated vehicle listings receive 80% positive feedback from sellers.

On benchmark performance: Claude Sonnet 4.5 scores 77.2% on SWE-bench Verified, the industry-standard test of AI ability to resolve real-world GitHub issues. That is the highest public score from any major AI coding assistant. Opus 4.6, the current frontier model, extends further on harder multi-step tasks.

The productivity ceiling has moved. Teams that used to plan around "one engineer, one feature, one sprint" are now shipping three features in the same window on the right classes of work.

Forrester's 2025 Total Economic Impact study of Claude for enterprise development reports 40-60% productivity improvements on suitable tasks, payback in under six months, and a 333% ROI over three years.

What happens to UK businesses that do not adopt Claude Code?

They fall behind. Quietly, then visibly. Their competitors ship features in days that used to take sprints. Their best engineers — the ones who care about tooling — leave for companies that let them use Claude Code. Technical debt compounds because nobody has time to refactor. The backlog grows. Sprint velocity plateaus. Hiring gets harder because the salary budget cannot keep up with the output gap.

This is not speculation. It is the pattern we see in UK SMEs that delayed AI adoption by twelve months in 2024-2025. The businesses that moved first did not just save time. They changed what was possible with the team they already had.

What does Claude Code look like inside a UK business?

It looks like a senior engineer who never takes a sick day, never asks for a raise, and reviews its own work before showing it to you. Three scenarios, composited from work Ampliflow has shipped across our UK client base.

Scenario 1 — A five-person UK SaaS team

A London-based SaaS company has five engineers and a growing product backlog. They adopt Claude Code first on one low-risk track: their test suite. A single engineer spends a Friday writing their CLAUDE.md — documenting the stack (Next.js 16, Convex, Stripe), the testing conventions (Vitest, Playwright for e2e, never mock the database), and the banned operations (no force-pushes, no skipping pre-commit hooks).

Within two weeks, Claude Code is writing tests for new features at the same pace as the engineers ship them. Coverage rises from 42% to 71%. Bug reports drop. By week six, the team is using Claude Code for small refactors and dependency upgrades — the maintenance work that used to slip between sprints.

The founder's metric: they shipped the equivalent of a sixth engineer's output without a sixth engineer's salary. That is £60,000-£80,000 of annual hiring cost that never materialised. The existing team reports that the work is more enjoyable, not less. The boring tickets go to Claude. The interesting ones stay with them.

Scenario 2 — A 40-person UK retail platform

A retail platform serving independent UK shops has 40 engineers split across web, mobile, and integrations. They run Claude Code in a structured pilot on the integrations team first — the work most often stuck behind third-party documentation.

The team wires MCP servers to their internal order-management system and their main partner APIs. An engineer who used to spend 2-3 days per partner integration now spends 2-3 hours, with Claude Code reading the partner's documentation, writing the adapter, generating tests against sandboxed endpoints, and producing the pull request. The engineer reviews, adjusts, and ships.

Governance is explicit. A shared CLAUDE.md lives at the repo root. Every PR generated by Claude Code is tagged. Security-sensitive areas are marked off-limits in the config. A weekly review reads the audit log.

After 90 days, the team's pull-request throughput is up 3.2×. The backlog of partner integrations has been cleared for the first time in eighteen months.

Scenario 3 — A non-developer UK operations lead

Not every Claude Code user is a developer. A UK operations lead running a growing professional-services business starts using Claude Code to automate reporting. Weekly revenue dashboards, client activity summaries, data reconciliation between a CRM and an accounting system — the work that used to require either a part-time analyst or several hours every Monday morning.

With Claude Code and a few carefully written skill definitions, the ops lead now runs a "Monday reports" command that reads from three systems, reconciles the discrepancies, flags anomalies, and produces a PDF report. What used to take four hours takes eight minutes. The reports are more accurate, because the reconciliation catches errors that the human eye missed.

Four hours of Monday-morning dread, replaced by eight minutes and a command. That is what Claude Code looks like when it reaches beyond the engineering team.

How does Ampliflow use Claude Code in production?

Through a methodology called Amplex — agentic orchestration with specialist harnesses and reviewing-model verification. Claude Code on its own is powerful. Deployed inside a disciplined production framework, it is transformative. The Amplex framework has four parts.

Specialist harnesses. Rather than run one generalist Claude Code instance across a whole codebase, we configure specialist harnesses — each with its own scoped CLAUDE.md, its own curated toolchain, and its own review rubric. A harness for the widget code looks different from a harness for the billing integration. This narrows the agent's context, reduces drift, and increases reliability.

Agentic orchestration. Harnesses compose. A single feature shipped through Amplex might involve the front-end harness, the integration harness, and the test-generation harness working in coordinated sequence, each handing off to the next with explicit state passed between them. This is how a team of one senior developer can supervise work equivalent to a team of four.

Reviewing-model verification. Every non-trivial change passes through an AI code review step before it reaches the human. A second model — typically Opus 4.6 — reads the proposed change with fresh context and grades it against a rubric: correctness, security, performance, style. The human reviews the reviewer's summary first, then the diff. Most regressions are caught here, before they ever reach staging.

Guardrails. Typed schemas on every inter-agent interface, circuit breakers on tool calls, entitlement gating on sensitive operations, and full audit logging. The framework defaults to refusing rather than guessing when evidence is thin.

Two of the systems we have shipped with Amplex are useful references. Cellbot is a Claude-powered widget and ops platform for UK repair businesses — widget chat at 153ms p95 latency, 71.82 requests per second sustained on a 20-concurrent-user load, a seven-stage conversion funnel live in production. SofaFlow is an enterprise-grade SaaS platform for UK furniture retailers — catalogue, visualisation, and order flow shipped with Claude Code and Amplex harnesses throughout.

We are publishing the deeper technical write-up of the specialist harness pattern on Promptology, our research publication. That paper covers the formal semantics, measured failure-rate comparisons, and the rubric design behind reviewing-model verification. Follow Promptology if you run engineering teams at scale and want the methodology behind the methodology.

What does Claude Code cost, and what is the ROI?

£17-£85/month per user, with payback typically in the first month for individual engineers. Forrester's six-month payback figure includes enterprise-wide rollout and onboarding; for a single developer on the Pro plan, payback is measured in days. Claude Code is priced per user. The current plans (as of April 2026) are:

Those are the sticker prices. The useful calculation is ROI.

A mid-level UK software engineer costs, fully loaded, about £75,000 per year. That is roughly £300 per working day, or £37.50 per working hour.

If Claude Code saves that engineer six hours per week, that is £225 of weekly value recovered for a £4 weekly subscription. Payback is immediate.

The maths at scale: 10 engineers × 6 hours saved × £37.50/hour × 52 weeks × 3 years = £351,000 of recovered engineering capacity. Total subscription cost: under £2,500. Forrester's 40-60% productivity figure suggests six hours per week is the floor, not the ceiling.

We are building an interactive calculator that runs the maths for any UK team size and effective rate — it goes live this month. In the meantime, book an audit and we will produce the numbers against your specific team.

The ROI case is the easy part. The hard part is doing the rollout properly.

The most common mistake we see in UK teams adopting Claude Code: they install it, skip the CLAUDE.md, skip the pilot, and ask it to "fix the whole app." Claude Code without context is like handing a senior contractor the keys to your office without telling them where the server room is or what the house rules are. Talented. Will still break something.

The teams that see 10% of the gains treat Claude Code as autocomplete. The teams that see 50-60% treat it as a new team member who needs a proper onboarding.

Write the CLAUDE.md. Scope the first project. Review every diff for the first fortnight. When Claude gets something wrong, add the rule to the CLAUDE.md — what the Anthropic team calls "every mistake becomes a constraint." The context file grows. The mistakes stop repeating. Then expand.

What do you need to start using Claude Code?

A Git repository, a CLAUDE.md file, and a review process. That is it. The starting checklist is shorter than you think.

1. A codebase that fits in a repository. Monorepos work. Polyrepos work. Greenfield projects work. What does not work is "my code lives on six developers' laptops and a few old Jenkins servers." The first job is usually some structural tidying.
2. A `CLAUDE.md` written in plain English. The single most important file in your Claude Code deployment. It tells the agent what your project is, what conventions to follow, which commands are forbidden, and where the important context lives. Most of the gap between good and great Claude Code usage is the quality of the CLAUDE.md. (See the example below.)
3. A review process that a human owns. Claude Code is an amplifier. It makes good teams faster and bad teams more productively wrong. A pull-request review standard — who approves what, what automated checks must pass, what cannot be bypassed — is the single biggest predictor of whether the rollout succeeds.
4. A small, scoped first pilot. The teams that succeed start with one low-risk track: test coverage, documentation, a specific refactor, a set of internal tools. They measure. They adjust the CLAUDE.md. They expand deliberately. The teams that fail try to do everything at once.
5. An honest view on governance. For SMEs, the Pro plan on a shared process is often enough. For regulated UK industries — financial services, legal, healthcare — the Enterprise plan with SSO, audit logging, and access controls is non-negotiable. A dedicated enterprise governance piece follows in this series.

What does a real CLAUDE.md look like?

Here is an example, simplified from a production UK SaaS:

`markdown

Project

Next.js 16 App Router + Convex + Stripe. UK B2B SaaS. Dev server: npm run dev (port 3001)

Rules

• British English throughout (colour, organisation, behaviour)
• Never run taskkill /IM node.exe (kills other dev processes)
• Never commit .env files or API keys
• Always run npm test before claiming a task is done
• Prefer editing existing files over creating new ones

Stack conventions

• Path alias: @/ maps to src/
• UI components: src/components/ui/ (shadcn)
• API routes: src/app/api/
• Tests: Vitest + Playwright. Never mock the database.

Deployment

• Push to main triggers Vercel auto-deploy
• Max 3 pushes per session (Vercel build-minute budget)

`

Ten lines of plain English. No special syntax. Claude Code reads this on every task and follows it like a brief.

"The instructions you give the agent — CLAUDE.md, skill files, context files — are more valuable than the code the agent writes."

— Boris Cherny, Engineering Lead, Claude Code

Context compounds. Code is disposable. The CLAUDE.md is the asset. We are releasing a free UK-stack CLAUDE.md generator later this month — subscribe to the newsletter to hear when it lands.

Frequently asked questions

Is Claude Code only for developers?

The best results come from developer-led teams, but non-developers use it successfully for data work, reporting, and internal automation — particularly once a technical colleague has written a good CLAUDE.md and set up the right skills.

How does Claude Code handle sensitive UK business data?

Claude Code operates only on the data you point it at. Anthropic does not train on API data by default and offers a zero-retention API option. The Enterprise plan adds contractual controls: configurable data retention, audit logging, and regional routing. For UK businesses handling regulated data under GDPR or sector-specific rules (FCA, SRA), the Enterprise plan is the right starting point.

Can Claude Code replace a junior developer?

It can do the work a junior developer typically does in their first six months — boilerplate code, tests, documentation, dependency upgrades, routine refactors. What it does not do is grow into a senior developer. The long-term case for hiring juniors is professional development, not cost. If you hire juniors to learn and contribute long-term, Claude Code does not change that. If you hire juniors to write boilerplate cheaply, Claude Code changes that significantly.

How long does it take to see ROI on Claude Code?

For teams that do a structured rollout, ROI is typically visible within the first 30 days and compounds from there. Forrester reports average payback under six months across the enterprises they studied. In practice, the bottleneck is almost never the technology — it is the rollout quality.

What is the difference between Claude Code and the Anthropic API?

The Anthropic API is the underlying model access. Claude Code is a specific product built on the API — a command-line tool with memory, tool use, agentic execution, and a defined developer experience. Most UK businesses do not need to build on the API directly. Claude Code covers the common cases and compounds with every improvement Anthropic ships.

Is Claude Code free?

Claude Code requires an Anthropic subscription. The Pro plan starts at £17/month per user and includes enough usage for moderate daily work. There is no permanent free tier, but Anthropic occasionally offers trial access. For teams evaluating whether it is worth the cost, the ROI calculation above shows payback within the first month for most UK engineering teams. The Max plan at £85/month per user is better suited to senior developers running long tasks or parallel sessions.

Is Claude better than ChatGPT for coding?

On the SWE-bench Verified benchmark — the industry standard for testing AI on real-world GitHub issues — Claude Sonnet 4.5 scores 77.2%, the highest public result from any major AI coding tool. ChatGPT (GPT-4o) is competitive on general coding questions, but Claude Code's terminal-native design, whole-codebase awareness, and CLAUDE.md memory system give it a structural advantage for production engineering work. For ad-hoc coding questions in a browser, both are strong. For shipping tested features across a real codebase, Claude Code is the best AI coding tool available today.

What if Claude Code does not work for our codebase?

It works on any codebase that fits in a Git repository. The question is not compatibility — it is configuration. A well-written CLAUDE.md and a scoped first pilot will surface whether Claude Code fits your workflow within the first week. If it does not, you have lost a few hours, not a few months. That is why we recommend starting with a single track, not a whole-team rollout.

What should you do next?

If you run a UK business with a small engineering team, the fastest way to find out what Claude Code could do for you is a free audit. We assess your stack, team size, and highest-impact opportunities, and return a specific Claude Code rollout plan within 48 hours. No obligation. No sales pitch. Just a plan you can execute with or without us.

Book a free Claude Code audit →

If you run a larger UK organisation — 50-plus staff, regulated industry, enterprise governance requirements — book a Technical Discovery Call. Forty-five minutes, free, no commitment. We cover architecture, delivery framework, security and governance, and commercial terms. You leave with a plan scoped to your environment.

Book a Technical Discovery Call →

The UK teams that move on Claude Code in 2026 will set the pace. The teams that wait will spend 2027 trying to close the gap.

Ampliflow is a UK AI frontier lab and growth agency based in Solihull, West Midlands. We ship production AI systems for UK SMEs and enterprises using Claude Code, the Amplex orchestration framework, and reviewing-model verification. Our case studies are named, our methodology is published, and our team builds with Claude Code daily.