Claude Code is the agentic AI coding assistant from Anthropic, the company behind the Claude AI models. It runs in your terminal, IDE, desktop app, browser and phone, searches and reads the relevant files in your codebase, edits files across a project, runs commands and tests, and asks for approval where your settings require it. Anthropic released it as a research preview in February 2025 and made it generally available in May 2025. In February 2026 Anthropic reported a Claude Code run-rate revenue of over 2.5 billion dollars, and cited a SemiAnalysis estimate that Claude Code authored about 4% of all public GitHub commits. This guide covers what Claude Code is, its features in 2026, how to install it, what it costs, how it compares to GitHub Copilot and Cursor, whether it is GDPR compliant, and when you should not use it.
Claude Code at a glance
| Topic | Status (checked 17 September 2026) |
|---|---|
| Developer | Anthropic |
| Type | Agentic AI coding assistant |
| Where it runs | Terminal (CLI), VS Code, JetBrains, desktop app, web (claude.ai/code), iOS and Android, Slack, GitHub Actions, GitLab CI/CD |
| Models | Claude Opus 5, Sonnet 5, Haiku 4.5 and Fable 5.1 |
| Context window | Up to 1 million tokens, depending on model and plan (Haiku 4.5: 200k) |
| Price | From 20 dollars per month (Pro), or pay per token via the API |
| Install | Native installer for macOS, Linux and Windows (commands below) |
| Certifications (Anthropic) | SOC 2 Type II, ISO 27001:2022, ISO/IEC 42001:2023, HIPAA-ready configuration |
What is Claude Code and how does it work?
Claude Code is an agentic AI tool: instead of suggesting the next line, it takes a goal (“fix the failing login test”), plans the steps, reads the relevant files, makes changes, runs the tests and reports back. Developers use it to build features, fix bugs, write tests, refactor code and automate Git and DevOps work. It runs on your own machine, or as a cloud session on Anthropic’s infrastructure or a self-hosted environment, and sends prompts and relevant code to a Claude model over an encrypted connection. It is available on several surfaces:
Terminal (CLI): the original and most complete interface.
IDE: native extensions for VS Code (also works in Cursor) and JetBrains.
Desktop app: a Code tab in the Claude app for macOS, Windows and Linux (beta).
Web and mobile: since October 2025, cloud sessions at claude.ai/code and in the Claude iOS and Android apps, running in isolated sandboxes.
Automation: Slack, GitHub Actions, GitLab CI/CD, automated code review and the Claude Agent SDK for building your own agents. Platforms such as Coder can also run Claude Code inside remote development workspaces.
Why Claude Code matters to modern development teams
Autocomplete tools help per line of code, but they do not plan work, run tests or fix what breaks. Claude Code does: it reasons across a whole repository and executes multi-step workflows.
The market has moved accordingly. Google’s 2025 DORA report found that 90% of the software professionals surveyed use AI, and over 80% say it has made them more productive. Google describes 65% as relying heavily on AI: 37% a moderate amount, 20% a lot and 8% a great deal. Trust lags behind: 30% of respondents trust AI output only a little or not at all.
Claude Code features in 2026
The capabilities that matter most in 2026 (availability can differ per surface and plan):
| Feature | What it does |
|---|---|
| Plan mode | Claude researches and proposes a plan before touching any file |
| Permission modes and auto mode | Control what runs without asking. In auto mode a classifier reviews actions instead of you; it is the built-in starting mode on Pro, Max and Team, and admins can change it |
| Checkpoints and rewind | Roll back code and conversation to an earlier point |
| Subagents and agent teams | Delegate subtasks to specialised agents with their own context |
| Skills | Reusable instructions and scripts Claude loads when a task calls for them |
| Plugins and marketplaces | Share bundles of commands, subagents, hooks and MCP servers across a team (since October 2025) |
| Hooks | Run your own scripts before or after actions, for example linting or blocking risky commands |
| MCP integrations | Connect Claude Code to ticketing, databases, logs, design tools and internal systems |
| Sandboxing | Isolate the file system and network for commands Claude runs, which Anthropic reports reduced permission prompts by 84% in its own usage |
| Background tasks and routines | Run long jobs in the background or on a schedule in the cloud |
| CLAUDE.md and memory | Project instructions and conventions Claude reads at the start of each session |
How to use Claude Code: getting started
Installation and initial setup
The native installer is the recommended method. It needs no Node.js and updates itself automatically.
# macOS, Linux and WSL
curl -fsSL https://claude.ai/install.sh | bash
# Windows PowerShell
irm https://claude.ai/install.ps1 | iexAlternatives are Homebrew (brew install --cask claude-code), WinGet (winget install Anthropic.ClaudeCode) and npm (npm install -g @anthropic-ai/claude-code, which now requires Node.js 22 or newer). Homebrew and WinGet installs do not update automatically. It runs on macOS 13+, Windows 10 1809+ and common Linux distributions with at least 4 GB of RAM.
You need a paid Claude plan (Pro, Max, Team or Enterprise) or a Claude Console account with API credits; the free plan does not include Claude Code. On first run, claude opens a browser-based login and stores the credentials locally, so align this with your organisation’s credential management policy.
Making your first Claude Code request
# Go to your project and start a session
cd /path/to/your/project
claude
# Ask in plain language
> Analyze this React component and identify performance bottlenecksClaude reads the relevant files, proposes changes as diffs and applies them within the limits of your permission mode. Start with low-risk questions such as “explain the structure of this project” before larger changes. Running /init creates a CLAUDE.md file with your project’s conventions, which noticeably improves results.
Deployment options: can you run Claude Code on-premise?
No, not fully. Claude Code itself can run on your own machines, on Anthropic’s cloud or in a self-hosted environment, but the Claude models are only available through Anthropic, a supported cloud provider or an LLM gateway in front of those. There is no on-premise or air-gapped model version. The closest options:
| Option | Where the model runs | Best for |
|---|---|---|
| Claude plans (Pro, Max, Team, Enterprise) | Anthropic | Individuals and most teams |
| Claude Console (API) | Anthropic | Pay-per-use, automation |
| Amazon Bedrock / Claude Platform on AWS | AWS | Organisations standardised on AWS |
| Google Cloud’s Agent Platform (successor to Vertex AI) | Google Cloud | Organisations on Google Cloud |
| Microsoft Foundry | Azure | Organisations on Microsoft Azure |
| Self-hosted environments (public beta, Team and Enterprise) | Cloud sessions run on your compute; inference only via the Anthropic API; not available with zero data retention | Keeping checkouts, builds and secrets on your own infrastructure |
For the regular CLI, infrastructure requirements are light: a supported OS, 4 GB RAM, network access to your model provider, and optionally an LLM gateway or proxy for central control. If you need models that run entirely on your own hardware, look at self-hosted open-weight LLMs instead.
Claude Code vs GitHub Copilot vs Cursor vs Codex
All major tools now offer agents; the differences are in approach and ecosystem.
| Claude Code | GitHub Copilot | Cursor | OpenAI Codex | |
|---|---|---|---|---|
| Starting point | Terminal-first agent, also IDE, desktop and web | IDE assistant with agent mode, cloud agent and CLI | AI-native code editor (VS Code fork) with cloud agents | CLI, IDE extension, cloud and desktop agent |
| Models | Claude only | Choice of Anthropic, OpenAI, Google and xAI models | Choice of frontier models plus own models | OpenAI GPT models |
| Typical fit | Multi-step tasks across a codebase, terminal workflows, custom agent setups | Teams standardised on GitHub that want model choice | Developers who want an AI-first editor | Teams already on ChatGPT |
| Entry price (individual) | 20 dollars per month (Pro) | 10 dollars per month (Pro), free tier available | 20 dollars per month (Pro), free tier available | Included in all ChatGPT plans, including Free and Go; Plus 20 dollars per month |
| Business price | Team from 25 dollars per seat per month | Business 19, Enterprise 39 dollars per seat | Teams from 40 dollars per user per month | ChatGPT Business, 25 dollars per user per month (20 billed annually) |
Monthly list prices, checked on the vendors’ pricing pages on 17 September 2026. Google moved Google AI Pro, Ultra and free users from Gemini CLI to Antigravity CLI in June 2026 (enterprise licences and paid API keys keep Gemini CLI). If you need a fully on-premise or air-gapped assistant, Tabnine offers that from 39 dollars per user per month on an annual subscription.
How to choose: pick Claude Code if your work involves larger autonomous tasks, terminal workflows or custom agent setups. Pick GitHub Copilot if your organisation already runs on GitHub and Microsoft and wants model choice under one contract. Pick Cursor if developers mainly want an AI-first editor. Many teams combine two: Claude Code also runs inside Cursor and VS Code.
When to use Claude Code
Software development and engineering teams
Claude Code is strongest on work that spans many files: implementing features end to end, refactoring, writing and fixing tests, migrating frameworks and investigating bugs from a stack trace. It also handles Git well, from commits and branches to merge conflicts and pull requests.
Data science and data engineering
Data teams use it to turn notebook experiments into production pipelines, write SQL and dbt models, and build quick internal visualisation apps without deep front-end knowledge.
DevOps and infrastructure automation
Infrastructure teams use Claude Code for configuration, CI/CD pipelines, Terraform and Kubernetes troubleshooting. Because it can run commands and read logs, it is effective at diagnosing failures, provided you restrict what it may execute in production environments.
Business process automation and non-technical use cases
Claude Code is not limited to developers. Novo Nordisk built NovoScribe, a documentation platform, with Claude Code. In its customer case the company reports that clinical documentation work went from more than 10 weeks to 10 minutes and clinical study report writing time fell by 90%. See our overview of Claude Code use cases for more examples.
Downsides of Claude Code
Claude Code is not the right tool for every situation:
- Unfamiliar or unusual architectures. Results are less consistent with complex event-driven systems or frameworks that are poorly represented in training data.
- Context drift in long sessions. Even with a context window of up to 1 million tokens, quality drops when sessions get long and noisy. Clear or compact context at logical boundaries; good context engineering helps.
- Usage limits. Subscription plans work with a 5-hour usage window plus weekly limits, and Anthropic has adjusted these several times. Heavy users can hit them mid-task.
- Review burden. Claude can produce a lot of code quickly. Without solid tests and code review, you trade writing time for debugging time.
- No offline or on-premise model. If code may never leave your network, Claude Code is not an option.
Security and GDPR considerations for enterprise deployment
Is Claude Code GDPR compliant?
Claude Code can be used in a GDPR-compliant way, but compliance depends on your configuration, contracts, data flows and use case, not on the tool alone. According to Anthropic’s data usage documentation:
Commercial plans (Team, Enterprise and API): Anthropic does not train models on your code or prompts, and Claude Code data is retained for 30 days by default. Anthropic’s data processing addendum is part of its Commercial Terms. Zero data retention is available to qualifying Enterprise customers.
Consumer plans (Pro and Max): data is used for training only if that setting is on. Retention is 30 days with the setting off and 5 years with it on. These plans do not come with the commercial terms, so we recommend a commercial plan for company code that may contain personal data.
Amazon Bedrock, Google Cloud and Microsoft Foundry: your agreement and region settings with that provider apply.
Beyond that, limit which repositories may be processed, keep secrets and unnecessary personal data out of prompts, and run a DPIA where required. This section is practical guidance, not legal advice. Anthropic holds SOC 2 Type II, ISO 27001:2022 and ISO/IEC 42001:2023 certifications and offers a HIPAA-ready configuration, but certifications do not replace your own assessment.
Known vulnerabilities
Like any tool that executes commands, Claude Code has had security issues. Examples of disclosed vulnerabilities and the version that fixed them (checked 17 September 2026; see Anthropic’s security advisories for the full list):
| CVE | Issue | Fixed in |
|---|---|---|
| CVE-2025-54794 | Path restriction bypass | v0.2.111 |
| CVE-2025-54795 | Command injection bypassing the approval prompt | v1.0.20 |
| CVE-2025-59536 | Project code could run before the startup trust dialog was accepted | v1.0.111 |
| CVE-2026-21852 | API key leak via a malicious repository setting before the trust prompt | v2.0.65 |
| CVE-2026-39861 | Sandbox escape via symlinks | v2.1.64 |
| CVE-2026-55607 | Sandbox escape via git worktree path confusion | v2.1.163 |
The pattern is consistent: the risk sits in untrusted repositories and project configuration. Keep Claude Code updated, only open trusted repositories, enable sandboxing, and use managed settings to enforce permission rules across your organisation. Read more about AI guardrails.
Cost management and usage optimization
Claude Code plans and pricing
| Plan | Price | Best for |
|---|---|---|
| Pro | 20 dollars per month (17 dollars billed annually) | Individual developers, moderate use |
| Max 5x | 100 dollars per month | Daily professional use |
| Max 20x | 200 dollars per month | Heavy, all-day use |
| Team Standard | 25 dollars per seat per month (20 dollars billed annually) | Teams, Claude Code included |
| Team Premium | 125 dollars per seat per month (100 dollars billed annually) | Teams with heavy Claude Code use |
| Enterprise | 20 dollars per seat, billed annually, plus usage at API rates | Large organisations, SSO, audit logs, admin controls |
API token pricing structure
| Model | Input (per million tokens) | Output (per million tokens) | Context window |
|---|---|---|---|
| Claude Fable 5.1 | 10 dollars | 50 dollars | 1M |
| Claude Opus 5 | 5 dollars | 25 dollars | 1M |
| Claude Sonnet 5 | 2 dollars | 10 dollars | 1M |
| Claude Haiku 4.5 | 1 dollar | 5 dollars | 200k |
API prices checked 17 September 2026. Current models no longer charge a long-context surcharge. Prompt caching reduces costs considerably for repeated context.
Usage management best practices
- Track consumption with /usage and set spend limits in the Console or admin settings.
- Use Sonnet or Haiku for routine work and reserve Opus or Fable for complex tasks.
- Clear context between unrelated tasks to save tokens.
- Plan intensive work around the 5-hour and weekly limits, or use API billing for automated workloads.
Implementation steps and best practices for organizations
Integrating Claude Code with development workflows
Work on a separate Git branch for every significant change, let Claude create a plan first in plan mode, and review that plan before execution. Keep a CLAUDE.md in each repository with conventions, commands and do’s and don’ts. Let Claude write tests first where possible, so it can verify its own work. In CI, the GitHub Actions and GitLab integrations let Claude review pull requests or pick up issues automatically.
Enterprise deployment strategies
Start with a pilot on non-critical projects: codebase Q&A, small bug fixes and isolated features. Configure managed settings centrally (allowed tools, permission rules, sandboxing), connect MCP servers for tickets, logs and documentation, and share proven workflows as skills or plugins. Plan for training as well: agentic coding asks for new skills, such as writing clear task specifications and reviewing larger changes.
The future of AI-assisted development
Claude Code roadmap: recent developments and direction
Based on recent releases, the direction is clear: longer autonomous work, more surfaces and more control for organisations. In twelve months Claude Code gained web and mobile sessions, plugins, skills, agent teams, auto mode, routines and self-hosted environments, and each model generation extended how long Claude can work without help. The developer’s role keeps shifting from writing code to specifying, reviewing and orchestrating agents.
AI skills become a hiring requirement
Gartner predicts that by 2027, 75% of hiring processes will include certifications and testing for workplace AI proficiency. Teams that benefit most start from concrete business problems, invest in training and build in security from day one.
How DataNorth can help
Implementing Claude Code takes more than an installation. Our Claude Code Workshoptrains your developers to work effectively and safely with agentic coding. An AI Assessment identifies where AI adds the most value in your organisation, our Live Demos show teams what is possible, and our AI Development and Implementation team builds solutions tailored to your technical challenges.
Frequently asked questions
What is Claude Code?
Claude Code is Anthropic’s agentic AI coding assistant. It works in the terminal, VS Code, JetBrains, the Claude desktop app, the browser and mobile. It searches and reads the relevant files in your codebase, edits files, runs commands and tests, and completes multi-step programming tasks based on instructions in plain language.
Is Claude Code free?
No. Claude Code requires a paid Claude plan (Pro from 20 dollars per month, Max, Team or Enterprise) or API credits through the Claude Console. The free Claude plan does not include Claude Code.
Which models can Claude Code use?
Claude Code works with Anthropic’s current models: Opus 5, Sonnet 5, Haiku 4.5 and Fable 5.1. According to Anthropic’s documentation (September 2026), Pro and Team Standard default to Sonnet 5, and Max, Team Premium, Enterprise and the API to Opus 5. Defaults can differ per provider and admin settings, and you can switch with the /modelcommand.
Is Claude Code better than GitHub Copilot?
It depends on the work. Claude Code suits multi-step tasks across a codebase and terminal workflows. GitHub Copilot is cheaper to start, offers several model providers and integrates deeply with GitHub. Many teams use both.
Does Claude Code run locally or on Anthropic’s servers?
Both. In the terminal or IDE, Claude Code runs on your own machine; cloud sessions run on Anthropic’s infrastructure or in a self-hosted environment. In every case, prompts and the code Claude needs are sent to the model through Anthropic, a supported cloud provider or an LLM gateway. The models themselves do not run on your hardware, so a fully offline setup is not possible.
Is Claude Code GDPR compliant?
It can be used in a GDPR-compliant way, but that depends on your setup and contracts. On Anthropic’s commercial plans (Team, Enterprise and API), Anthropic does not train on your data, retains Claude Code data for 30 days by default and includes a data processing addendum in its Commercial Terms. Via AWS, Google Cloud or Azure, that provider’s terms apply. For company code that may contain personal data, we recommend a commercial plan over Pro or Max.
