Claude Code is an AI coding platform launched by Anthropic in 2024 and matured in 2026. Unlike completion-style tools (GitHub Copilot) or chat-window-style tools (Claude.ai), Claude Code lets AI agents directly operate your terminal, file system, Git, and external APIs to complete multi-step engineering tasks. Starting in April 2026, Claude Code enters a “systematized” phase— a redesigned desktop version, cloud Routines scheduling, and the .claude folder becoming the standard structure for permissions and skills. This article covers everything end to end, from installation, interface, and permissions to hands-on practice.
Claude Code Core Positioning
Claude Code’s design philosophy is “AI agents need a controllable workspace,” rather than putting AI into an existing IDE. It consists of three components:
CLI command-line tool: cross-platform and installable on macOS / Linux / Windows, with the official native interface
Desktop application: fully redesigned on April 14, 2026, supports multiple sessions side by side, and integrates the terminal with file editing
Cloud Routines: launched as a research preview on April 14, 2026, supports scheduling, API-triggering, and GitHub Webhook event-driven execution
All three share the same .claude configuration structure, allow interchangeable models, and use MCP to work with the same external tools. This makes Claude Code a complete AI engineer workstation, rather than a single-point tool.
Model choice: Opus 4.7 / Sonnet 4.6 / Haiku 4.5
Claude Code defaults to using Sonnet 4.6 as the main model, but you can switch to stronger Opus 4.7 or faster Haiku 4.5 via
/model
:
Model Suitable scenarios Notes Opus 4.7 Long tasks, complex refactors, multi-step planning New tokenizer increases 37–47%, so costs need to be re-evaluated Sonnet 4.6 Daily coding, medium-complexity workflows (default) Balanced performance and cost; Claude Code defaults to this Haiku 4.5 Large-scale file scanning, batch classification, summarization Haiku 3 was retired on 4/19; switch to 4.5
CLI Quick Installation
The fastest way to get started:
curl -fsSL | bash claude
On the first launch, it will prompt you to log in (using your Claude.ai account or an Anthropic API key), and to read the current directory’s
.claude/
configuration. After that, just run claude in your project directory to enter an interactive REPL.
For Linux / Windows users who can’t access curl, you can install via npm or pip:
npm: npm install -g @anthropic-ai/claude-code
pip: pip install claude-code
.claude folder: the control hub for permissions, commands, and skills
The .claude folder is Claude Code’s root directory for configuration. The structure is as follows:
CLAUDE.md
: the project-level “system prompt”; Claude Code reads it every time it starts. Write team rules, architecture overviews, and disabled items so the agent follows organizational conventions
commands/
: custom slash commands (e.g., /deploy, /review) that can be reused across projects
skills/
: skill files for specific tasks; the agent automatically calls them when it detects related intent (e.g., image processing, API integration scripts)
agents/
: sub-agent definitions, used to break down complex tasks into multiple specialized agents
hooks/
: event-trigger scripts (tool_use, stop, prompt_submit, etc.), which can be used to do logging, automated testing, and security checks
settings.json
: a permissions list (allow / prohibit which Bash commands, MCP servers), model selection, and environment variables
The permissions architecture is divided into three layers: allow (auto-grant), ask (ask every time), and deny (directly block). This design lets AI agents operate autonomously within a bounded sandbox, preventing accidental actions.
MCP: standard interface for external tools
Claude Code connects to external services via MCP (Model Context Protocol). The official analogy is that “MCP is AI’s USB-C”—any tool that implements an MCP server can be called by Claude Code.
Common MCP servers include: file systems, Git, GitHub, Slack, Notion, Figma, Postgres, Playwright (browser automation), and more. After registering an MCP server in
.claude/settings.json
, Claude Code will automatically add the tool’s capabilities to its tool list.
In March 2026, MCP’s monthly downloads exceeded 97 million, becoming a real-world standard for agentic applications.
Routines: AI engineers for cloud scheduling
Claude Code Routines was launched as a research preview on April 14, 2026, turning Claude Code from “you need to keep your computer on to run it” into “a cloud service that stays running.” There are three trigger methods:
Schedule: runs at fixed times (e.g., daily 06:00 to clean up expired issues)
API: starts via external system HTTP calls (integrates with Datadog, PagerDuty, CI)
GitHub Webhook: automatically triggers on events such as PR opened, push, or issue creation
Routines are executed in Anthropic’s cloud, so users don’t need to keep their computer powered on. Pricing is model token fees plus $0.08 per runtime hour; a 45-second routine is about $0.001 runtime plus token fees. The number of daily runs depends on the subscription tier: Pro 5 times, Max 15 times, Team / Enterprise 25 times.
Desktop version: an AI workstation with multiple sessions side by side
Claude Code’s desktop version was fully redesigned on April 14, 2026, supporting macOS and Windows. New version features:
Multiple sessions side by side: run multiple agents in parallel to handle different projects
Terminal integration: run shell commands inside the application and view Claude Code output
File editing: built-in Monaco (the same core editor as VS Code)
Dispatch mobile remote control: start and monitor desktop sessions from a phone app
The desktop version shares the .claude configuration used by the CLI, and the two can switch seamlessly. For users accustomed to IDE workflows, the desktop version is a more user-friendly entry point than a pure CLI.
Claude Code vs Cursor vs GitHub Copilot
Project Claude Code Cursor GitHub Copilot Positioning AI engineer platform (CLI + desktop + cloud) AI-native IDE IDE code completion AI Engineer Platform (CLI + desktop + cloud) Core strengths Long tasks, multi-step planning, cloud Routines IDE experience, Tab completion, inline refactoring, and deep GitHub ecosystem integration Model Opus 4.7 / Sonnet 4.6 / Haiku 4.5 Multi-model (Claude / GPT / Gemini) Multi-model (GPT-5.x / Claude / Gemini) Monthly plan Pro $20 / Max $100、$200 Pro $20、Business $40 Pro $10、Business $19
Selection recommendation: If you’re used to IDE workflows, choose Cursor; if your team is GitHub-centered, choose Copilot; if you need long tasks, multi-step automation, and cloud scheduling, choose Claude Code. It’s also common for engineers to use all three together in 2026.
Claude Code vs Claude Cowork
Claude Cowork is another product from Anthropic, positioned as an “AI desktop collaborator that doesn’t touch the API”—users don’t need commands, there’s no .claude configuration, and they can directly use natural language to operate Claude and carry out desktop work (writing emails, organizing spreadsheets, automating web actions).
Simple distinction: Cowork = an AI assistant for non-engineers; Claude Code = an AI agent platform for engineers. The two can be used at the same time, sharing the Claude.ai account and billing allowances.
Billing and plans
Claude Code is billed via a Claude.ai subscription or API token charges:
Pro ($20/month): Sonnet 4.6 main, with Opus 4.7 usage limits, 5 daily Routines
Max ($100/month): higher Opus 4.7 allowance, 15 daily Routines
Max ($200/month): highest Opus 4.7 allowance, suitable for full-time AI workers
Team / Enterprise: custom allowances, SSO, RBAC, 25 daily Routines
API: billed per token (Opus 4.7 $5/$25 per M token; note that the new tokenizer makes actual consumption +37–47%)
Important updates in April 2026
4/14: desktop version fully redesigned
4/14: Routines research preview launched
4/15: enterprise edition moves first to usage-based billing
4/16: Opus 4.7 launches; Claude Code supports switching
4/17: Claude Design launches, enabling one-click implementation in Claude Code
4/19: Haiku 3 retired; Opus 4.7 tokenizer cost changes appear
Frequently Asked Questions (FAQ)
Is Claude Code free?
Claude Code itself is free software, but you need a Claude.ai subscription (minimum $20/month Pro) or an API key to use the model. The official side does not offer a standalone purchase plan for Claude Code.
How is Claude Code different from Claude.ai’s web version?
Claude.ai is a chat interface: you can only input text and view Claude’s replies. Claude Code, however, can directly operate the files on your computer, run shell commands, use MCP tools, and complete multi-step real work.
Is Claude Code safe? Could it delete the wrong files?
By default, it is designed as “ask before executing any new action.” Advanced users can customize the permissions list in
.claude/settings.json
to allow / ask / prohibit which commands. The official recommendation is to use Git version control for important projects in combination and avoid granting Claude Code write access to the entire home directory.
What’s the difference between Routines and local cron?
Local cron requires your computer to be on to run; Routines run in Anthropic’s cloud, operate anytime, support GitHub Webhook and API triggers, and are driven by Claude Code rather than you writing shell scripts.
Can Claude Code be used in Taiwan?
Yes. Users in Taiwan can register Claude.ai normally, install the Claude Code CLI and desktop version, use the API and Routines, and there are no regional restrictions. Payment methods support credit cards and enterprise contracts.
Can I migrate my Cursor or Copilot usage habits?
Yes. Claude Code can write organizational rules in .claude/CLAUDE.md (equivalent to Cursor Rules or Copilot Instructions), and supports connecting MCP to services such as GitHub, Slack, Notion. The recommended first step is to copy the contents of your existing .cursorrules or copilot-instructions.md into CLAUDE.md to get started quickly.
What is Claude Code in this article? The most complete 2026 guide: full breakdown of CLI installation, desktop version, Routines automation, MCP, and the .claude permissions structure first appeared on Lianxin ABMedia.
