Getting set up.

Here's the good news first. You don't need to know how to code, you don't need a developer on staff to "set it up," and you don't need a fancy project management tool. Claude writes the code, sets up its own computer, and keeps track of the work in its own notes. What you need is a Mac, a Claude plan, a handful of accounts, and one afternoon. This guide is Mac only, so if your office runs on Windows, that's the one thing to change first.

The computers.#

Your laptop is where you work. You talk to Claude there, review what it made, and approve anything risky. A MacBook Air or Pro with Apple silicon is right, and memory is the one spec worth paying for: get 24 GB or more. 16 GB works, but it chokes when Claude runs several jobs at once, and Staying in charge has the story of what that looks like. Your phone is the one you already have. It's for checking in, approving things, and recording voice notes that turn into instructions.

Later you'll probably want a Mac mini as an always-on worker: Apple silicon, 16 GB or more, 512 GB of storage, plugged into ethernet and a battery backup if you can. It runs scheduled jobs overnight and long tasks while your laptop is closed, and it holds a second copy of everything. Don't buy it on day one. Add it once you notice you're leaving your laptop open overnight so a job can finish. That's the sign, and The always-on setup covers it.

The Claude plan.#

Claude Code, the part of the app that does real work on your computer, comes with every paid plan. The free plan doesn't include it. The chart below lines up the monthly prices as of September 2026.

If more than one person in your business will use Claude, read the right plan for a business before you buy. The short version: a Team seat costs a little more than Pro and keeps your data out of model training by default.

Pro fits trying this out or light daily use. Max 5x fits daily use as your main way of working, and Max 20x fits several sessions all day plus overnight jobs on the Mac mini. Your usage resets on a rolling five-hour window, with weekly limits on top, so move up when you keep hitting the limit. The Mac mini signs in to the same account, so if overnight jobs start eating your daytime usage, move up a tier or give the business Team seats. People sometimes ask about running a second personal Max account instead. Check Anthropic's Consumer Terms before you try that; Team seats are the cleaner answer for a business.

One setting deserves a minute of your attention. Pro and Max run under consumer terms, which include a choice about whether your chats help train future models. You'll find it under Settings, Privacy, "Help improve our AI models." If you handle client data, set it on purpose rather than by accident, or use Team or Enterprise, which run under commercial terms.

The other accounts, and the real cost.#

Set these up in your company's name, all signed in with one work email that you control. Every tool hangs off that one identity, so when you hire someone later you give them access to specific things and never your login. You'll want Google Workspace for email, calendar, Drive and Meet (about $15 to $25 a user; meeting recordings and transcripts need Business Standard or higher, which Working with Claude every day explains). You'll want GitHub, where every project lives, set up as an organization for the company rather than a personal account ($0 to $4 a user). Vercel puts your websites and tools online the moment Claude saves them to GitHub, for about $20 a month. Then a password manager like 1Password (about $8 a user), which holds every human password plus the one master key for your machine secrets (Teaching Claude your business explains those). Tailscale, free for personal use, gives your laptop a private path to the Mac mini from anywhere, and Jump Desktop (about $35, once) lets you see the mini's screen, though the Mac's built-in Screen Sharing works too.

There are two ways to pay for Claude, and it helps to keep them straight. A plan is a flat monthly price for using Claude in its own apps, and this whole guide is built on it. The API is different: you pay per token (a small chunk of text, roughly three quarters of a word) through the Anthropic Console, and a plan doesn't include any API credit. The API is for software you build that calls Claude on its own, like a script that writes 1,500 product descriptions overnight. Anthropic's own figures put API-billed Claude Code at roughly $150 to $250 per developer per month on average, and a runaway script can cost far more. So when a script really does need the API, Claude uses the cheapest model that does the job and gives you a dollar estimate before it spends anything.

Your first afternoon.#

The old way to set up a work Mac is a forty-step checklist full of terminal commands. Skip it. On your first afternoon you handle only what a human has to do, then hand Claude one long request and spend most of the afternoon watching it work. The timeline below shows roughly how the hours split.

Turn the Mac on and make your user account. Pick a short, one-word, lowercase user name, like tim. It ends up in thousands of file paths, and spaces cause headaches. Sign in with your Apple ID, and when macOS offers to turn on FileVault, say yes. FileVault encrypts the disk, so a stolen laptop is just a brick. Save the recovery key in your password manager, not on a sticky note. Then open System Settings, go to General, then Software Update, and install everything, restarting as many times as it asks.

Next, install Google Chrome, make it your default browser, and sign in with your work Google account. Install your password manager and its Chrome extension while you're there. Every human password lives in it, and only in it.

Now the main event. In Chrome, go to claude.com/download and download the app for Mac. Open the file, drag Claude into your Applications folder, and open it from there. Sign in with the account that has your paid plan, and you'll see three tabs across the top: Chat, Cowork and Code. From here on, the app keeps itself up to date. While you're at it, add the Claude in Chrome extension from the Chrome Web Store, sign in with the same account, and pin it to the toolbar.

Last, click the Code tab. When it asks for a folder, make a new one called Code inside your home folder (the one named after you) and choose it. You'll see it written as ~/Code, where the ~ is shorthand for your home folder, and every project will live there. Leave the permission mode, the setting for how often Claude checks with you, on Manual for now, so Claude asks before each step.

Handing Claude the rest.#

Now you tell Claude who you are, what your company and GitHub organization are called, and what you want the machine to become. The full request is long, but what it builds is simple, and the stack below shows it from the bottom up. First come the basic tools: Apple's command line tools, Homebrew (the installer most Mac developers use for everything else), git, which tracks every change to a file, and gh, the small command GitHub uses to sign in and create projects. Then Claude signs in to GitHub and Vercel, with you clicking through the login pages. Then it builds its memory, a private brain of notes kept in a folder under ~/.claude, interviews you with a few short questions to fill it in, and sets up a small job that backs the brain up to GitHub every minute. Last, it writes your standing orders into ~/.claude/CLAUDE.md, the one file it reads at the start of every session.

Set up this Mac as my work computer. I'm not technical, so you run every command yourself. Install the basic tools, sign me in to GitHub, build your memory, and keep it backed up. Only stop for things that truly need my hands, like a password or a popup.

You press Enter, and the afternoon becomes Claude's. Expect the handful of popups from the timeline, each one a click or a login, and read every one before you click. In Manual mode, Claude also asks before each command with a one-line description of what it's about to do; if it says what you'd expect, approve it. The Homebrew installer is a normal Mac package that opens in Chrome and asks for your Mac password, and the GitHub sign-in asks for a short code Claude shows you in the chat. The popups that deserve a real pause are the ones from macOS asking whether an app can access a folder or control your computer. Access to your Code folder is expected. Full Disk Access, Accessibility and Screen Recording are bigger keys to hand over, so approve those only when you know which step needs them, and ask Claude why if you don't.

At least once, Claude will stop and hand you a command to run yourself. It's a habit, not a necessity. Reply "You run it." If it says it can't, reply "Figure out another way, and tell me the one thing only I can do." That's your first taste of the pushback habit from Staying in charge.

Once "you run it, I only do logins" is a rule, sign-ins and consent clicks are all you ever do by hand.

When Claude says it's finished, don't take its word for it. Ask it to prove the setup works: the version of each tool it installed, the latest backup of the brain on GitHub and how many seconds ago it happened, and the list of notes in the brain. Then close the app, reopen it, start a new session in the Code tab on your Code folder, and ask, "What do you know about me?" If it answers from the brain, with your name, your business and your rules, day one is done. It's worth opening the standing orders file once, too. The first lines look something like the excerpt below, in plain English.

# How we work
- I'm the boss, you're the crew. Do the whole job yourself.
- Never hand me a command to run. Only stop for logins and popups.
- Prove it before you say done: a screenshot, a test, a live link.

## Never without my clear yes
- Send any email or message (draft it instead)
- Spend money or change a budget
- Delete anything

Three tabs, one job per session.#

With the machine set up, it's worth ten minutes to learn your way around the app. It's more like learning where the light switches are in a new office than learning software. Chat is more than talk: it searches the web, uses your connected apps, and makes files. Cowork is Anthropic's tab for knowledge work that isn't about code, like documents, research, a report or a list of tasks, and it's a comfortable place to hand off a job that doesn't need your projects or Claude's memory. Code is where Claude works directly on your computer, inside a folder you pick. Ignore the name. You'll use it for websites, spreadsheets, emails, research and bookkeeping, and it's the tab that can see your projects, your brain and your scheduled jobs. That's why this guide lives there.

Cowork gets its own chapter later in this book, Cowork, for the rest of the team, because it's where most of your staff will spend their time.

Each conversation in the Code tab is a session, and the easiest way to think about one is as a person you handed a single job to. "Build the new pricing page" is a session. "Answer the client's email" is another. Four or five running at once is normal, and the sidebar shows which are working and which are waiting on you. Rename them as you go ("Acme homepage redesign," not "New session 14") so you can find them later. Sessions opened on the same folder share its files, so for work inside a single project, turn on the worktree option, which gives each session its own copy so they don't step on each other. Long sessions summarize their older parts on their own and keep going. Each session also has a model picker and an effort setting: use the most capable model for real work, turn effort up for hard problems and down for quick questions. Heavier models use up your plan limits faster, and fast mode, which speeds up the top model, bills as extra usage beyond your plan.

How often Claude asks first.#

Before Claude runs a command or edits a file, it can check with you, or it can just go ahead. You choose with the mode picker next to the message box. Start in Manual, which asks before every edit and command, for your first week, while you learn what normal looks like. There's also Accept edits, which changes files without asking but still checks before other commands, and Plan, which researches and writes a plan and changes nothing until you approve it. Plan is great for big jobs: "Plan the new website, then show me." The scale below puts all five modes in order.

The goal is Auto. In Auto, Claude does the work without stopping to ask, while a background safety check blocks risky actions. Once you notice you're approving everything without reading it, and your guardrails are written down, that's when to switch. Your safety comes from the rules in the brain, like "never send an email, only draft it," plus Auto's safety check, not from clicking Approve 200 times a day. To make Manual less tedious in the meantime, tell Claude to add the harmless commands you keep approving to its allowed list, and it will stop asking about them.

The fewer interruptions, the more work gets done. Earn your way to fewer interruptions.

Bypass permissions, the last stop on the scale, hides behind a separate setting for a reason. For a business owner, the answer is never. Mode names do change as the app updates, but the idea stays the same.

Your Chrome, your skills, and your phone.#

Claude can use the web in two ways. The built-in browser is a clean window inside the app with no saved logins, good for previewing a site Claude is building. Claude in Chrome, the extension you installed on day one, lets Claude work in your own Chrome with all your logins already there: Google Ads, Search Console, your CRM, your domain registrar. It pauses and hands back to you at login screens and CAPTCHAs. Tell Claude once which one you want, and save it as a rule: "Always use my Chrome for web tasks unless I say otherwise."

Skills are saved how-tos. Each one is a folder of instructions Claude picks up on its own when a task matches, like your brand rules, your month-end bookkeeping, or how you onboard a new client. When you catch yourself explaining the same process twice, say "turn this into a skill." Plugins are bundles of skills and connections other people have built for sales, finance, design and so on, so install what fits and ignore the rest. A session can also spin up helpers, called subagents, to do parts of a job at the same time: "Research these ten competitors at once, then summarize."

Finally, your phone. With Remote Control, you can pick up a live session from the Claude phone app to check progress, answer a question Claude is waiting on, or give it the next job. Type /remote-control in a session, or turn it on for every new session in Settings under Claude Code. The computer has to stay on and awake. You may also hear about Dispatch, in the Cowork tab, which lets you message Claude from your phone to start brand-new work on your computer. As of September 2026 it's a limited beta that isn't open to new users, so treat it as a bonus if you have it. Pair Remote Control with voice notes, and a lot of real work gets started from the passenger seat.