Claude Code in business: the guide to getting started and managing it
Guide published on October 12, 2026, compiled from official Claude Code documentation and Anthropic offers consulted that day.
Claude Code is Anthropic's code agent: a task is described to him in everyday language, he reads the project, modifies the files, launches the commands and tests, and then proposes the result. In business, the question is no longer whether it works, but who uses it, with which offering, on which projects, and under what rules. This guide answers these four questions, for developers as well as for business teams who want to build their own tools.
What Claude Code actually does
A typical code assistant completes a line or answers a question. Claude Code works on an entire project: it explores files to understand how the code is organized, proposes a plan, writes changes to multiple files, runs tests, fixes what breaks, and then prepares a commit or pull request. Anthropic's documentation describes it as a tool that "reads your codebase, modifies files, executes commands, and integrates with your development tools."
Common uses include: developing a feature, fixing a bug based on an error message, writing missing tests, documenting legacy code, migrating a library, automating pull request reviews, or scheduling recurring tasks. It belongs to the same family of tools as Codex at OpenAI, which we compare in Claude Code or Codex: which code agent to choose.
Where to use it
Claude Code runs on multiple surfaces that share the same engine: the CLAUDE.md instruction file, settings and MCP connectors of a project apply everywhere.
| Surface | For who |
|---|---|
| Terminal (command line) | Developers: the most complete version |
| VS Code and JetBrains extensions | Developers who want to stay with their publisher |
| Desktop application | Developers and business profiles: sessions visible side-by-side, without a terminal |
| Browser (claude.ai/code) and mobile application | Tasks launched on a hosted repository, which continue in the cloud even when the computer is turned off. |
| GitHub Actions, GitLab CI/CD, Slack, Chrome | Automatic review of pull requests, ticket sorting, web application debugging |
Which offer to choose
Claude Code does not offer a free version: a paid Claude subscription or an API account is required. As of October 12, 2026, Anthropic's prices in dollars are as follows.
| It offers | Price displayed | When to choose it |
|---|---|---|
| Pro | $20 per month ($17 annually) | Individual trial, one-time use |
| Max | $ 100 or $ 200 per month | Intensive use of a developer |
| Team | $25 per seat per month ($20 annually), premium seat $125 ($100 annually) | Team: centralized administration, settings imposed from the console |
| Enterprise | Quote only | Large organization: SSO, SCIM provisioning, compliance |
| API Console | For consumption | Scripted usage or volume billing |
| Amazon Bedrock, Google Cloud, Microsoft Foundry | According to the cloud contract | Organization that wants to inherit its cloud controls and billing |
Note from the documentation: Cloud sessions, scheduled routines, automatic code review, remote control, and the Chrome extension all require a claude.ai account. Organizations relying solely on an API key or their cloud provider will not have access to these features without a Team or Enterprise subscription. Anthropic recommends the Team or Enterprise plans by default, which combine Claude Code and the chatbot under a single subscription per seat.
Install it in five minutes
On macOS, Linux or WSL, only one command in the terminal:
curl -fsSL https://claude.ai/install.sh | bashOn Windows, in PowerShell:
irm https://claude.ai/install.ps1 | iexThen simply open a project folder and type claude The tool requires a login on first launch. Homebrew, WinGet, and Linux package managers are also supported, and the native installation updates automatically. On Windows, Anthropic recommends Git for Windows so the agent has a Bash shell. Business users can avoid using the terminal by using the desktop application or claude.ai/code.
The first day with Claude Code
Four reflexes make the difference between an impressive demonstration and reliable use.
- Write a file named CLAUDE.md At the root of the project: coding conventions, test commands, and what should never be touched. The agent reads it at each session, and it can also read an existing AGENTS.md file.
- Start in Plan mode The agent proposes their approach without modifying any files. The plan is validated, then it is allowed to execute.
- Choose the right authorization method Manual, where each action is validated; Acceptance of changes; or Auto, where a classifier allows safe actions to pass and blocks risky ones.
- Rely on checkpoints Each file modification is preceded by a backspace point. Actions on remote systems, such as a database or a deployment, are not undone in this way: they are governed by authorization rules.
And what if you're not a developer?
This is the most underestimated aspect. A management controller, a communications officer, or a quality manager can describe in French the tool they need and see it built: a script that consolidates Excel files, a small tracking application, a dashboard, a document converter. The desktop application and the web version are sufficient, without needing to open a terminal.
The key skill is no longer programming, but specifying clearly, verifying the result, and knowing what can be deployed in production. This is called vibe coding, and it's precisely where projects go off the rails without proper frameworks: a tool that works on three lines of test code but not on the actual file, sensitive data copied to a personal folder, an application that no one knows how to maintain. We detail this approach, and the framework that the IT department can endorse, in our Vibe coding training in companies.
How the IT department manages Claude Code
Claude Code enforces organizational policy through managed settings that override each developer's local configuration. These settings are distributed in four ways: from the claude.ai administration console (Team and Enterprise offerings), via a system policy distributed through the workstation management tool (plist on macOS, registry on Windows), via a file on the workstation, or as a last resort through the Windows user registry, which the documentation advises against using as a control method since it can be modified without administrator privileges.
| What the IT department can impose | How does |
|---|---|
| Allow, validate, or block tools and commands | Authorization rules (allow, ask, deny), which developers can add to but not remove |
| Preventing the circumvention of permissions | Disabling the mode that skips validations; authorization rules reserved for managed settings. |
| Choose the session start mode | Default authorization mode, or removal of Auto mode |
| Isolate the execution | System-level sandbox, with a whitelist of network domains |
| Give instructions to everyone | A CLAUDE.md organizational file loaded in each session, impossible to exclude |
| Limit connections | Restrictions on MCP servers and plugin sources |
| Track usage and costs | Analysis tables, OpenTelemetry telemetry, expense tracking |
On the data side, the Team and Enterprise offerings do not use code to train default models. To learn more about this, see our article. Claude's Safety and Compliance in Business.
Deploy without losing control: four steps
In most organizations, a few developers already use a code agent, often on a personal subscription. Deployment involves scaling up from these pioneers to a common practice.
- Take inventory : who uses which tool, on which projects, with which account.
- Set the frame : offer selected, settings managed, CLAUDE.md organization, list of projects open to the agent and those that remain closed.
- Train and appoint referents A practical session on a real-world case is better than a presentation. That's the principle of our Code Party, practical training in Claude Code and Codexwhere each participant delivers a tool on a case from their team.
- Measure the performance of their campaigns : time saved, pull request acceptance rate, incidents avoided, cost per seat. Without these figures, widespread adoption remains a matter of conviction.
For the entire process, from choosing the tools to measuring the gains, see our page AI coding: developing business applications with coding agentsand to understand what surrounds the model in these tools, What is an AI harness?
F.A.Q
Is Claude Code free?
No. You need a paid Claude subscription (Pro, Max, Team, or Enterprise) or a pay-as-you-go API account. The cheapest option is the Pro plan, at $20 per month.
Do you need to know how to program to use Claude Code?
Not for beginners: the desktop application and web version allow you to describe a need in French and review the result. But verifying what the agent produces, and deciding what can be used in production, requires a method and some basic knowledge.
Does Claude Code work on Windows?
Yes, natively in PowerShell or WSL. Anthropic recommends installing Git for Windows so the agent has a Bash shell; without it, it uses PowerShell.
Is my code used to train Anthropic models?
Not enabled by default on Team and Enterprise plans. On individual plans, data usage for training is configured in the account settings.
Can Claude Code be used via Azure or AWS?
Yes: Claude Code connects to Amazon Bedrock, Google Cloud, and Microsoft Foundry. However, some features, such as cloud sessions, routines, and automatic code review, require an additional claude.ai account.
Sources
- Anthropic, Claude Code documentation: presentation and installation, accessed on October 12, 2026.
- Anthropic, deploying Claude Code in an organization, accessed on October 12, 2026.
- Anthropic, Claude's offers, accessed on October 10, 2026.
- Anthropic, Use Claude Code with a Pro or Max plan.