Claude Code is Anthropic's coding agent. It runs in a terminal, reads the files in your project, writes code, runs commands, and works through multi-step jobs without you approving every step. You describe the outcome in plain language, and it goes and does the work.
I run my whole practice on it. This is what it is, how I actually use it, what it costs, and what I think a business owner should take from it.
What Claude Code actually is
It's a command line tool you install with npm. You open a terminal in a project folder, run claude, and talk to it. It reads the project structure, plans the work, edits files, runs commands, checks its own output, and keeps going until the job is done or it hits something it needs you for.
The thing that makes it different from other AI coding tools is that it has no interface of its own. No editor, no window, no tabs. That sounds like a limitation and it's the whole point. A tool with no screen can run on a machine with no screen. It can keep working on a Mac in a cupboard while you're in a meeting, and you can reach it over SSH from a phone on a train.
Tools built into an editor need you sitting at a computer with the editor open. This one doesn't, and everything interesting I do with it follows from that.
How I run it
My business runs on an AI agent operating system, and Claude Code is the engine inside it. The full write-up is in that post; here's the short version.
A Mac Mini at home stays on around the clock, headless, with agent sessions running inside tmux so they survive any disconnection. Claude Code does most of the work, with Hermes, Grok and Codex alongside for jobs that suit them better. I reach the machine over SSH from whatever device I have on me, including my phone through ShellDrop, an iOS terminal app I built for exactly this.
The part that does the heavy lifting is the memory, and it's plain Markdown. A CLAUDE.md file teaches the agent how the business works, who the clients are, how I write, and what has already been decided. Under that sits a folder per client and per project with its own state and history. Every session reads those files before it does anything, so I've stopped re-explaining my own business to a machine.
The agent also has an identity. It has an email address, it sends and it receives, and I can hand it a job from anywhere by emailing it like a colleague. Ask it for a landing page, a dashboard or a small internal app and it builds one. The pages on this site were built this way.
If you want that setup on your own machine with someone walking you through it, the half-day course builds an agent operating system for your business with you: half a day on Zoom, two seats at most, starting from opening Terminal for the first time.
Which model does the work
Claude Code runs on Anthropic's Claude models, and you choose which one is driving.
The Opus models are the ones I use for anything that needs judgement: architectural decisions, multi-file refactors, work where the agent has to hold a lot of context and make calls without me. They cost more and they're slower. When an agent is running a business, the smartest model available is the right seat to fill.
The Sonnet models are faster and cheaper, and they're the right fit when you already know exactly what you want and the job doesn't need much reasoning. Quick edits, straightforward features, repetitive work. The quality gap on hard problems is real, so I don't try to save money on the thinking.
Version numbers move every few months and any guide that leans on them goes stale fast. The useful shape to remember is that there's a heavier model for judgement and a faster one for volume, and the heavier one should be running anything you'd be annoyed to have done badly.
What it costs
Claude Code itself is free to install. What you pay for is the model doing the work, and there are two ways to do that.
- A Claude subscription. Claude Code is included with the paid Claude plans. Pro is US$20 a month billed monthly (less if you pay annually), and the Max plans start at US$100 a month for people working with it all day. Predictable bill, usage limits that reset.
- Anthropic's API, per token. You pay for what the model reads and writes on each job. No monthly commitment, and the bill scales with how hard you push it.
Prices move, and Anthropic's pricing page carries the current numbers. The way I weigh it is against my own hourly rate. If a subscription covers work that used to take a morning a week, the maths does itself. If you're reaching for it on things you'd finish faster by hand, you'll feel that in the bill.
What it's good at
Building things end to end is where it's most impressive. Ask for a feature and you get the front end, the backend logic, the database changes and the type definitions, because it reads the whole project before it starts. Restructuring is similarly strong: renaming a service and updating every file that touches it is a one-sentence instruction.
Debugging is where I've got the most value. Paste an error, and it reads the relevant files, traces the problem back, and often finds the cause three files away from where the error surfaced.
It handles chains of work well. "Build the page, run the build, screenshot it, fix the responsive issues, rebuild" is a single instruction, and it works through the steps, checking as it goes. That is what agentic AI looks like on an ordinary Tuesday.
And it's good at explaining code that already exists. "How does authentication work in this project" gets an accurate answer, which is worth a lot when the code in question is something you wrote six months ago.
Where it still needs a person
It can't see. It writes the CSS and has no idea whether the page looks right. Anything visual gets my eyes on it before it goes anywhere, every time.
Jobs where the success criteria are a feeling rather than a fact turn into loops. "Make it feel premium" costs more rounds than "use this palette, this spacing and this typeface". It does its best work when what "done" means is unambiguous.
On very large codebases it needs steering. Without instructions about which parts matter, it can spend time and money reading files it didn't need. The CLAUDE.md file is most of the fix.
And it can be confidently wrong. Long jobs get checked. Anything with client money or a client relationship attached gets read properly before it leaves my machine.
What a business owner should take from this
The part worth your attention is that the ordinary work can leave your desk and come back close to finished, with the whole business already loaded, and none of it requires you to be sitting at a keyboard while it happens.
A chat window in a browser starts cold every time you open it. Claude Code with a memory behind it knows your clients, your numbers, your tone and what you decided last quarter, and it can go and act on all of that.
Building that takes a day of setup and a working method that isn't obvious from the documentation. The Claude Code course builds it with you in half a day, on your machine, loaded with your business, with your first custom app running on top of it before we finish.
Working out where AI fits in your business?
AI advisory for founders and business owners who want a straight answer on tools and implementation.
AI advisory services →Getting started yourself
1. Get a Claude account. A Pro subscription is the simplest starting point, and it includes Claude Code. An API account with credits works too if you'd rather pay per job.
2. Install it. npm install -g @anthropic-ai/claude-code, which needs Node.js on your machine.
3. Point it at something real. You'll learn more in an afternoon on a project you care about than on a toy example. Navigate to the folder and run claude.
4. Ask for concrete things first. "Add a 404 page to this project" gives you a cleaner read on it than "make the site better". Broaden the instructions as you get a feel for what it handles well.
5. Write a CLAUDE.md file. Tell it about the project: the stack, the conventions, what matters, what not to touch. It's the biggest lever on output quality and it takes twenty minutes.
6. Put the heavier model on the hard jobs. The cost difference is small next to the time spent fixing work from a model that wasn't up to the task.