Before you start
- A Claude Pro, Max, Team or Enterprise subscription, or a Claude Console account (the Free plan doesn't include Claude Code)
- Windows 10 (1809) or later, macOS 13 or later, or a recent Linux distribution (Ubuntu 20.04+, Debian 10+)
- 4 GB of RAM and an internet connection
- Having already opened a terminal; Git installed, preferably
In plain words
Claude Code is Claude installed in your terminal, with permission to read your project, modify files and run commands. Instead of copying and pasting code between a conversation and your editor, you tell it "fix this bug", "add a test", "write the backup script", and it explores the project, makes the changes, runs the tests and shows you the result. It's a colleague working directly in your workshop: very effective, as long as you tell it clearly what it's allowed to touch.
Claude Code is as useful to developers as to administrators: PowerShell or Bash scripts, Ansible files, configurations, Dockerfiles, documentation for a Git repository. To get started with Claude in the app, see Getting started with Claude. Information checked on 20 October 2026 against the official documentation; commands change, and the documentation is the authority.
1. Where to use Claude Code
| Interface | For whom |
|---|---|
| Terminal (CLI) | The reference interface, described on this page |
| VS Code and JetBrains | Extensions that integrate Claude Code into the editor, with changes displayed |
| Desktop app | Claude Code without a terminal, with a graphical interface |
| Web (claude.ai/code) | Sessions in the cloud, on a GitHub repository, with nothing to install |
| GitHub Actions, GitLab CI/CD | Automating code reviews or fixes in continuous integration |
2. Installing
The native installer is the recommended method: it installs a standalone executable and updates it automatically. You don't need administrator rights.
macOS, Linux, WSL:
curl -fsSL https://claude.ai/install.sh | bash
Windows, in PowerShell:
irm https://claude.ai/install.ps1 | iex
Windows, in Command Prompt (CMD):
curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd
PowerShell or CMD?
The PowerShell prompt starts with PS C:\, the CMD prompt with just C:\. If you see The token '&&' is not a valid statement separator, you're in PowerShell; if you see 'irm' is not recognized, you're in CMD.
With a package manager
| System | Command | Updates |
|---|---|---|
| Windows | winget install Anthropic.ClaudeCode |
Manual: winget upgrade Anthropic.ClaudeCode |
| macOS | brew install --cask claude-code |
Manual: brew upgrade claude-code |
| Debian, Ubuntu, Fedora, RHEL, Alpine | Signed apt, dnf and apk repositories (see the setup documentation) | With system updates |
| All, with Node.js 22+ | npm install -g @anthropic-ai/claude-code |
npm install -g @anthropic-ai/claude-code@latest |
Never sudo npm install -g
The documentation explicitly forbids it: installing a global npm package as root creates permission problems and a security risk. Use the native installer instead.
Windows: native or WSL?
| Option | For whom | Note |
|---|---|---|
| Native Windows | Windows projects and tools (PowerShell, .NET) | Preferably install Git for Windows: Claude Code then uses Git Bash; otherwise, it uses PowerShell |
| WSL 2 | Linux projects, Bash scripts, Docker | Run the Linux installer inside WSL; it's also the only Windows option that supports the sandbox |
Checking the installation
Open a new terminal, then:
claude --version
2.1.xxx (Claude Code)
For a full diagnosis (installation, settings, updates):
claude doctor
If claude isn't recognised, the installation folder isn't in the PATH yet: close and reopen the terminal, then see the troubleshooting page in the documentation.
3. Signing in and starting a first session
Go to the project folder, then start Claude Code:
cd ~/projects/admin-scripts
claude
On first use, a sign-in page opens in the browser: sign in with your Claude account. The credentials are then kept. To switch account: type /login in the session.
Start by having it explore the project before asking for changes:
what does this project do? what's the folder structure and which tools does it use?
Then a first simple task:
add to the backup.sh script a check that the destination disk is mounted
before starting the copy, and a clear error message otherwise
Claude Code finds the file, proposes the change, applies it according to the permission mode, and can run the script or the tests to check.
4. Permission modes: who decides?
This is the most important setting. It sets what Claude Code can do without asking you. You switch from one mode to another with Shift+Tab during the session, or at launch with --permission-mode.
| Mode | What runs without asking | When to use it |
|---|---|---|
Manual (default) |
Reads only | Sensitive work, unfamiliar code, discovering the tool |
| acceptEdits | Reads, file edits, common file commands (mkdir, cp, mv…) |
Iterating on code that you review |
| plan | Reads only, then presents a plan to approve before any change | Exploring and preparing a significant change |
| auto | Almost everything, with safety checks made by a second model (classifier) | Long tasks on a well-understood project |
| dontAsk | Only explicitly allowed tools; the rest is refused | Locked-down scripts and continuous integration |
| bypassPermissions | Everything, with no checks at all | Only in an isolated container or VM |
The starting mode has changed
Since recent versions (2.1.283 and later), interactive sessions in the terminal and VS Code start in auto mode: Claude edits most files and runs most commands without asking you. For a first encounter or sensitive work, start with:
claude --permission-mode default
or switch to Manual mode with Shift+Tab at the start of the session.
5. The CLAUDE.md file: the project's memory
Each session starts from an empty context. The CLAUDE.md file holds the standing instructions that Claude Code reads at start-up: build and test commands, conventions, architecture, known pitfalls.
To generate a first version, run this in the session:
/init
Claude analyses the project and writes a CLAUDE.md, which you then complete with what it can't guess.
| Location | Scope | Shared with |
|---|---|---|
~/.claude/CLAUDE.md |
Your preferences, for all your projects | You alone |
./CLAUDE.md or ./.claude/CLAUDE.md |
The project | The team, through Git |
./CLAUDE.local.md |
Your personal settings for this project (add it to .gitignore) |
You alone |
File managed by the organisation (for example /etc/claude-code/CLAUDE.md on Linux) |
The whole organisation | All users |
Example for a repository of administration scripts:
# Administration scripts
- Bash scripts for Debian 13, PowerShell 7 for Windows Server 2025.
- Test a Bash script: `shellcheck script.sh`, then run it on the test VM `test01`.
- Never run a script on a production server.
- Every script starts with a help block (`--help`) and logs to /var/log/admin/.
- Variables in upper case, functions in lower case with underscores.
- User in examples: spo.
A CLAUDE.md can import others with the @path/file.md syntax. Claude Code also keeps an automatic memory of your corrections and preferences, specific to each repository.
Short and precise
The shorter and more concrete the instructions, the better they're followed. CLAUDE.md is an instruction, not a barrier: to forbid an action no matter what, use permission rules (part 7).
6. The essential commands
In the terminal
| Command | Effect |
|---|---|
claude |
Opens an interactive session in the current folder |
claude "task" |
Opens a session with a first request |
claude -p "question" |
Answers once, then returns control (useful in a script) |
claude -c |
Resumes the last conversation in this folder |
claude -r |
Chooses a past conversation to resume |
claude update |
Installs the latest version immediately |
claude doctor |
Diagnosis of the installation and settings |
In a session
| Command or key | Effect |
|---|---|
/help |
List of available commands; typing / also shows commands and skills |
/init |
Generates or improves the CLAUDE.md |
/clear |
Clears the conversation to start a new task |
/compact |
Summarises the conversation to free up space in the context |
/context |
Shows what's taking up the context (CLAUDE.md files loaded, etc.) |
/model |
Changes model |
/permissions |
Views and edits access rules |
/config |
Settings, including the update channel (latest or stable) |
/login |
Switches account |
| Shift+Tab | Changes permission mode |
| Esc | Interrupts Claude mid-action |
/exit or Ctrl+D twice |
Quits |
7. Securing your use
Permission rules
Rules add to the modes: allow (allow without asking), ask (always ask), deny (always refuse, even in bypassPermissions mode). They go in a settings file:
| File | Scope |
|---|---|
~/.claude/settings.json |
You, for all projects |
.claude/settings.json |
The project, shared through Git |
.claude/settings.local.json |
You, for this project only |
Example .claude/settings.json for an infrastructure repository:
{
"permissions": {
"allow": [
"Bash(shellcheck *)",
"Bash(git status)",
"Bash(git diff *)"
],
"ask": [
"Bash(git push *)",
"Bash(ssh *)"
],
"deny": [
"Read(./.env)",
"Read(./secrets/**)",
"Bash(rm -rf *)"
]
}
}
Good habits
- Work in Git: commit before a significant task. Every change can be reviewed with
git diffand undone easily. - No secrets in the project: passwords and keys go in a vault or environment variables, and
.envfiles are blocked from reading by adenyrule. - Never straight to production: have Claude Code work on a copy, a branch or a test VM, then deploy yourself.
- Turn on the sandbox (
/sandbox, on macOS, Linux and WSL 2): it limits what the commands it runs can reach on the disk and network. bypassPermissionsmode only in a disposable container or VM, with no access to your data or credentials.- Be wary of the content it reads: a file, a web page or a ticket can contain hidden instructions (prompt injection). Manual mode or
askrules keep a human in the loop for sensitive actions. - Review the code produced before running it on an important system, as you would a colleague's code.
8. Going further
| Feature | What it is |
|---|---|
| MCP | Connecting Claude Code to external tools (database, ticketing, monitoring) with claude mcp add |
| Skills | Reusable procedures that Claude applies when the task calls for them |
| Subagents | Specialised assistants (code review, tests) that Claude Code launches for part of the work |
| Hooks | Scripts triggered at certain moments (before a command, after an edit) to check or format automatically |
| Plugins | Ready-to-install bundles of skills, commands, hooks and MCP servers |
| Non-interactive mode | claude -p in a script or a scheduled task |
Anthropic offers a free course, Claude Code 101, on Claude Academy.
Key points
- One-command installation (
curl … | bashorirm … | iex), automatic updates; check withclaude --versionandclaude doctor. - Requires a Pro, Max, Team or Enterprise plan or a Console account.
- Shift+Tab changes the permission mode; recent sessions start in auto: switch to Manual for sensitive work.
/initcreates the CLAUDE.md: short, concrete project instructions.denyrules for secrets and dangerous commands, Git to review and undo everything, never straight to production.