What Is the Claude Code CLI and How Do You Use It?

What Is the Claude Code CLI and How Do You Use It?

Most AI coding tools live inside an editor. They watch what you type and offer the next few lines, which is useful but keeps them in the role of a very fast autocomplete.

The Claude Code CLI starts somewhere else. It runs in your terminal, reads your project, and takes on whole tasks rather than individual lines.

That difference shapes everything about how you install it and how you work with it. There is no extension to enable and no editor to configure before you start.

There is also a detail that trips up anyone following an older tutorial: the install command most guides still show is no longer the recommended one.

This guide covers what the CLI is, what it needs, how to install it on each platform, how updates work, and when the terminal is the wrong place to be.

The Short Answer

The Claude Code CLI is Anthropic's agentic coding tool for the command line. You install it with a one-line script, run claude inside a project directory, and describe what you want in plain language.

On macOS, Linux and WSL the recommended install is curl -fsSL https://claude.ai/install.sh | bash. On Windows PowerShell it is irm https://claude.ai/install.ps1 | iex.

It needs a paid Claude plan. Pro, Max, Team, Enterprise and Console accounts all include access, and the free claude.ai plan does not.

Your code is not the only thing that needs an audit. Get a free SEO audit and see how your site is performing in search.

What the Claude Code CLI Actually Is

Claude Code is an agent rather than a suggestion engine. You give it a task, it reads the files it needs, forms a plan, and makes changes across however many files the job touches.

The terminal is its native home, with extensions available for VS Code and JetBrains, plus the ability to run in the cloud and against GitHub repositories.

That architecture explains the kind of work it is good at. Multi-file refactors, migrations, and bugs you cannot locate are its strongest cases, because all three need something that can look around on its own.

What it needs on your machine

The requirements are modest. macOS 13 or later, Windows 10 1809 or later, or a recent Ubuntu, Debian or Alpine Linux will do.

Hardware-wise you need 4 GB of RAM and an x64 or ARM64 processor. A working internet connection is required, since the model runs remotely rather than on your machine.

Any of Bash, Zsh, PowerShell or CMD works as the shell. The search dependency, ripgrep, ships with Claude Code, so you do not normally install it separately.

What it needs from your account

Access is tied to a plan rather than to a separate purchase. A Pro, Max, Team, Enterprise or Console account covers it, and the free tier does not include it at all.

You can also point it at a third-party provider instead. Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry are all supported paths for teams whose models live there.

How Do You Install the Claude Code CLI?

The native installer is the recommended route on every platform. It downloads a self-contained binary rather than installing anything through a language runtime.

1. Open a terminal in any directory. The installer does not care where you run it from, since it installs the binary into your user directory.
2. Run the command for your system. On macOS, Linux or WSL that is curl -fsSL https://claude.ai/install.sh | bash. In PowerShell it is irm https://claude.ai/install.ps1 | iex.
3. Wait without feedback. The install command prints no progress while it downloads, which looks like a hang and is not. Give it a moment.
4. Open a new terminal window. The installer adds its directory to your PATH, and an already-open shell will not have picked that up.
5. Check it works. Run claude --version. A working installation prints a version number; a command not found error means the install directory is not on your PATH yet.
6. Run the diagnostics if anything looks off. claude doctor prints a read-only report on install health and settings problems without starting a session.
7. Start your first session. Move into a project directory and run claude. That opens an interactive session scoped to that project.
8. Log in. Follow the browser prompts the first time. If you have set an API key in your environment and approve it when asked, the login prompt is skipped.

If you would rather not use a package manager at all, that is the whole process. There is nothing else to configure before the first session.

Technical setup is one thing, visibility is another. Get a free SEO audit and find out what is holding your rankings back.

The package manager routes

Homebrew works on macOS and Linux with brew install --cask claude-code. There is a second cask, claude-code@latest, which tracks new releases as they ship rather than lagging by about a week.

Windows users can use WinGet with winget install Anthropic.ClaudeCode. Debian, Fedora, RHEL and Alpine users have signed apt, dnf and apk repositories available.

The trade-off is updates. Package manager installs do not update themselves by default, so you pick them up through whatever upgrade routine you already run.

Why the npm command in old tutorials is not the answer

The npm package still exists and still works. npm install -g @anthropic-ai/claude-code installs the same native binary that the standalone installer does.

It is no longer the recommended path, though, and the project's own repository labels installation via npm as deprecated in favour of the installers above.

If you do use it, you need Node.js 22 or later, and you should never run it with sudo. Permission errors on a global npm directory are a common and avoidable mess.

One nuance worth knowing: the installed binary does not run on Node at all. npm is only the delivery mechanism, which is part of why it stopped being the preferred one.

How Do Updates Work?

This depends entirely on how you installed it, and the difference is the single most common source of confusion after the install itself.

Native installations update themselves. Claude Code checks on startup and periodically while running, downloads in the background, and applies the new version the next time you start it.

Homebrew, WinGet, apt, dnf and apk installations do not auto-update. You run the relevant upgrade command yourself, or set an environment variable to have Claude Code run it for you.

To force an update immediately, run claude update. It reports the version it moved you to, or tells you that you are already current.

There are two release channels. The default, latest, gives you new features as they ship; stable runs roughly a week behind and skips releases with serious regressions.

You set the channel through /config in a session or in your settings file. A newly launched model sometimes needs a version newer than stable serves, which is the main reason to stay on latest.

When the Terminal Is the Right Surface

The CLI is the right choice when the work is agentic. Handing over a task that spans several files is what it was built for, and the terminal keeps that interaction simple.

It is also the right choice for anything you want to script or run in CI, since a command-line tool composes with everything else in that environment.

If you would rather not work in a terminal at all, there is a desktop application that gives you the same tool with a graphical interface.

The editor extensions sit in between. They suit people who want Claude Code's agentic behaviour without leaving the editor they already have open.

A faster development workflow does not help if nobody finds your site. Get a free SEO audit and fix the gaps.

Limitations Worth Knowing

Windows has a split worth understanding before you commit. Native Windows works fine, but sandboxed command execution is only supported under WSL 2.

On native Windows, installing Git for Windows is optional but useful. Without it, Claude Code runs shell commands through PowerShell instead of Bash.

Alpine and other musl-based distributions need a few packages installed first. Bash and curl for the install itself, plus libgcc, libstdc++ and ripgrep at runtime.

Network storage is a genuine trap. A running session reads from the binary on disk, so an in-place upgrade on a shared home directory can crash sessions on other machines.

There is no offline mode. The model runs remotely, so a dropped connection stops the session rather than degrading it.

And the account requirement is absolute. There is no free tier for Claude Code, so evaluating it means starting a paid plan or using a third-party provider.

Claude Code CLI vs the Alternatives

Against editor-first tools, the split is directness. An autocomplete-style assistant is always suggesting; Claude Code waits to be asked and then does considerably more.

Against other terminal agents, the differentiator is how much of the codebase it reads before acting. Planning across files is the behaviour it is designed around.

Against the desktop app and the editor extensions, it is the same tool with a different front end. Pick on how you prefer to work rather than on capability.

For teams, the deciding factor is usually governance. Managed settings let an organisation enforce a release channel and a minimum version across everyone.

Final Thoughts

Install with the native script, verify with claude --version, and ignore any tutorial that leads with the npm command. That covers almost every setup problem people hit.

After that, the thing worth learning next is how to package up the workflows you build. Our guide to how Claude Code plugins work covers bundling skills, agents and hooks into one install.

Run claude doctor whenever something behaves oddly, and remember that the release channel you are on decides whether a brand new model is available to you yet.

Shipping faster is only half the job. The other half is making sure the pages you ship can actually be found, which is a different kind of audit entirely.

See what your site looks like to search engines today. Get a free SEO audit and get a prioritised list of fixes.