What Is Claude Code? Anthropic’s AI Coding Agent Explained

Claude Code is Anthropic's agentic coding tool: a terminal-based (and now IDE, desktop, and browser-based) AI assistant that reads your codebase, edits files, runs commands, and ships pull requests on its own, rather than just answering questions in a chat window. It launched in February 2025 alongside Claude 3.7 Sonnet as a research preview and has since expanded into a full product line that includes the Claude Code CLI, IDE extensions, a desktop app, a web version at claude.ai/code, and a developer-facing Agent SDK.

What Claude Code Actually Is

Anthropic's own documentation describes Claude Code as "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools," available in the terminal, an IDE, a desktop app, and a browser. That phrase — "agentic" — is the key difference from a chatbot. Instead of waiting for you to copy and paste code snippets back and forth, Claude Code operates directly on your project: it searches your files to understand structure, proposes and makes edits, executes shell commands, runs your test suite, and can commit changes and open pull requests against GitHub or GitLab.

The tool first shipped on February 24, 2025, in Anthropic's Claude 3.7 Sonnet announcement, which called it "a command line tool for agentic coding" that let developers "delegate substantial engineering tasks to Claude directly from their terminal." Anthropic framed early Claude Code as "an active collaborator" that stays in the loop with the developer at every step, rather than a black box that silently rewrites a codebase.

How Claude Code Works

Under the hood, Claude Code runs an agentic loop: you give it a task in plain language, and the model plans an approach, calls tools (file read/write, search, Bash commands, web fetches) to carry it out, checks its own work, and reports back. Anthropic's product page puts it simply — Claude Code "builds the plan, asks clarifying questions, and handles work that runs for hours or days," asking permission before it changes files or runs commands.

That agent loop is available across several surfaces that all share the same underlying engine, settings, and connected tools:

• Terminal CLI — the original, full-featured interface; you run claude inside any project directory.
• IDE extensions — official plugins for VS Code (and VS Code–based editors like Cursor) and JetBrains IDEs, adding inline diffs and plan review directly in the editor.
• Desktop app — a standalone application with a dedicated "Code" tab for reviewing diffs visually, running several sessions side by side, and scheduling recurring tasks.
• Web and mobile — claude.ai/code runs Claude Code in the browser with no local setup, useful for kicking off long tasks on repos you don't have locally; sessions can be picked up later on the Claude iOS/Android app.

Because every surface points at the same engine, a project's CLAUDE.md instructions, settings, and connected MCP servers carry over whether you start a task from a terminal, an IDE, or your phone.

Getting Started: Install and First Run

Installing Claude Code takes one command. Anthropic's current documentation recommends the native installer over older package-manager routes:

curl -fsSL https://claude.ai/install.sh | bash

irm https://claude.ai/install.ps1 | iex

cd your-project
claude

Claude Code is also installable via Homebrew (brew install --cask claude-code), WinGet, or Linux package managers including apt, dnf, and apk. Once it's running, you can type a task directly — "write tests for the auth module, run them, and fix any failures" — and Claude Code will plan the work, make the edits, and run the tests itself. The CLI also works non-interactively for scripting and CI, for example piping log output into claude -p to triage errors automatically. Anthropic's full Claude Code documentation covers advanced setup, troubleshooting, and every supported surface in detail.

Plan Mode, Auto Mode, and Permission Controls

One of the most-asked questions about Claude Code is what "plan mode" and "auto mode" actually do, because both control how much Claude is allowed to do without asking first. Claude Code organizes this through permission modes, which you cycle through in the terminal with Shift+Tab or set with the --permission-mode flag:

• Default (Manual) mode — Claude can read files freely but must ask before writing, editing, or running most commands. This is the most conservative mode, suited to sensitive work.
• Plan mode — Claude explores the codebase and proposes a plan using only read-only tools, and blocks any edits until you explicitly approve that plan. It's built for scoping out a change before anything touches your files.
• Accept Edits mode — file edits and common filesystem commands run without a prompt, while Claude still asks before anything riskier.
• Auto mode — a second model, described by Anthropic as a classifier, reviews each tool call before it runs and approves safe actions automatically, blocking things like mass file deletion or sensitive data exfiltration. Anthropic calls it "a safer long-running alternative to --dangerously-skip-permissions," and notes it reduces risk "but doesn't eliminate it entirely" — isolated environments are still recommended for unattended runs.
• Bypass Permissions mode — skips essentially all prompts; Anthropic restricts this to isolated containers or VMs.

Per Anthropic's permission modes documentation, auto mode became the built-in starting mode for interactive terminal and VS Code sessions with Claude Code version 2.1.283 and later (it was previously limited to Pro, Max, and Team plans), and reached general availability for all users on July 10, 2026. Organizations can disable it centrally through managed settings if they want every session to start in manual review mode instead.

Hooks, MCP, and Subagents

Three features give Claude Code its depth beyond a single chat-and-edit loop:

Hooks are user-defined handlers — shell commands, HTTP calls, MCP tool calls, or even subagents — that fire automatically at points in Claude Code's lifecycle, such as PreToolUse, PostToolUse, SessionStart, or Stop. According to Anthropic's hooks documentation, a common pattern is a PreToolUse hook that blocks destructive Bash commands like rm -rf before they execute, or a PostToolUse hook that automatically runs a linter after every file edit. Hooks are configured in JSON settings files and can be scoped to a single project, your whole machine, or an entire organization.

MCP, the Model Context Protocol, is how Claude Code reaches outside your codebase to tools like Jira, Google Drive, Slack, Postgres, or internal company systems. You add a server with a single command — for example, claude mcp add --transport http notion https://mcp.notion.com/mcp — and once connected, Claude can read and act on that system directly instead of working only from text you paste into the chat. Anthropic maintains full setup instructions in its MCP documentation. Teams that want a packaged way to share MCP servers, hooks, and skills across a company typically turn to Anthropic's plugin and marketplace system, which bundles these extensions for easy install.

Subagents let Claude Code delegate a task to a specialized assistant that runs in its own context window, with its own system prompt, tool access, and permissions, returning only a summary to the main conversation. Anthropic documents several built-in subagents, including Explore (fast, read-only codebase search), Plan (read-only research during plan mode), and General-purpose (multi-step tasks needing both exploration and edits). Developers can also write custom subagents as a Markdown file with YAML frontmatter, stored in a project's .claude/agents/ folder, restricting each one to specific tools or even a cheaper model like Haiku. By default, subagents can nest up to three layers deep, and a session can run up to twenty subagents concurrently.

The Claude Agent SDK

For teams that want to build their own agent rather than use the CLI interactively, Anthropic offers the Claude Agent SDK, available for Python and TypeScript. Anthropic describes it as giving developers "the same tools, agent loop, and context management that power Claude Code," packaged as a library you embed in your own application. The SDK carries over Claude Code's built-in tools, hooks, subagents, MCP support, permission system, sessions, skills, and plugins — the same capabilities documented for the CLI, just programmable. It sits alongside, not in place of, the Claude Code CLI (built for interactive, day-to-day use) and the lower-level Claude API/Client SDK (where you write the tool loop yourself).

Pricing: Subscription vs. API Access

Claude Code is bundled into Claude's consumer and business subscription plans rather than sold as a separate product, though you can also pay for it purely through API usage. Per Anthropic's own pricing page, Claude Code is not available on the free tier:

On a subscription plan, Claude Code usage shares the same rolling five-hour usage pool as chatting with Claude on the web, desktop, or mobile app, with paid plans adding weekly limits on top. If you'd rather not deal with session limits at all, you can skip the subscription model entirely and authenticate Claude Code with an ANTHROPIC_API_KEY, paying per token through an Anthropic Console account instead — a common choice for CI pipelines, automated workflows, or teams that want granular, metered billing rather than a flat seat price. When subscription limits are reached mid-task, paid plans can also opt to continue at standard API rates rather than wait for the limit to reset.

Claude Code vs. Claude.ai Chat vs. Claude Cowork

It's easy to confuse Claude Code with other Claude products, since they all run on the same underlying models. The practical distinction is what each one is allowed to touch:

• Claude.ai chat is a conversational interface — you talk to Claude in a browser, desktop, or mobile app, optionally using Projects and Artifacts, but it does not have standing access to your local filesystem or terminal. (Claude.ai does offer a separate GitHub integration for pulling repository context into chat, but that's distinct from the Claude Code CLI.)
• Claude Code is built specifically for software engineering: it operates directly on a codebase, running in a terminal, IDE, desktop app, or browser, with tools to read and write files, run shell commands, execute tests, and manage git branches, commits, and pull requests.
• Claude Cowork uses the same underlying agentic architecture as Claude Code but is aimed at non-coding knowledge work rather than software projects. It runs inside the Claude Desktop app without opening a terminal, works across local files and connected apps like Slack or Google Drive, and — per Anthropic's support documentation — can run tasks in the cloud, in an isolated environment, rather than only on your own machine. Pandromeda has covered how Claude Cowork now runs in the cloud by default on Pro and Max in more detail.

In short: reach for Claude.ai chat for conversation and research, Claude Code for anything that touches a repository, and Claude Cowork for broader office-style work that spans files and apps but isn't code.

Who Claude Code Is Actually For

Anthropic pitches Claude Code at individual developers and engineering teams handling the kind of work that's tedious but necessary: writing tests for untested code, fixing lint errors across a project, resolving merge conflicts, updating dependencies, and writing release notes, in addition to building new features and fixing bugs described in plain language. It also connects to CI/CD: Anthropic documents GitHub Actions and GitLab CI/CD integrations for automated code review and issue triage, and a Slack integration that turns a bug report mentioned to @Claude in a channel into a pull request. For teams comparing Claude Code against other AI coding tools before committing to one, Pandromeda's GitHub Copilot vs. Claude Code comparison goes deeper on how the two stack up feature-for-feature.

Because the underlying engine is also shipped as the Agent SDK, some of what looks like "Claude Code" in the wild is actually a custom-built agent running on the same technology — for instance, teams packaging repeatable internal workflows as reusable Claude Skills that Claude Code or an SDK-based agent can load automatically.

What to Do Next

If you write or maintain code and haven't tried Claude Code yet, the lowest-friction way in is the terminal install above, run inside a real project with a Pro or higher subscription already active. Start in plan mode on anything you're unsure about, let auto mode handle routine long-running tasks once you trust its judgment on your codebase, and only reach for hooks, MCP servers, and custom subagents once the default workflow feels too limiting. For larger engineering orgs, the next step is usually evaluating Team or Enterprise seats against straight API billing, and deciding whether a custom agent built on the Agent SDK makes more sense than the interactive CLI for automated, headless use cases. Anthropic keeps its documentation at code.claude.com current as the product evolves, making it the first place to check before any workflow built around Claude Code's permission modes, hooks, or subagents.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage