What Is Claude Code? Anthropic's AI Coding Tool

Claude Code is Anthropic's agentic coding tool, announced alongside Claude 3.7 Sonnet in a limited research preview in February 2025 and later made generally available. Unlike GitHub Copilot or Cursor, which live inside an editor, Claude Code is terminal-first: you run claude in a project directory and converse with an agent that can read your codebase, write and edit files, run shell commands, search the web, and plan multi-step work. In Anthropic's own early testing, the tool completed tasks in a single pass that would normally take engineers 45 minutes or more of manual work.

It is Anthropic's flagship product for software engineering and the canonical example of what people now call vibe coding — describing an outcome in plain language and letting an agent handle the keystrokes. This guide covers how it actually works under the hood, what changed as the underlying models moved from Claude 3.7 Sonnet through Opus 4.5, 4.7, and 4.8, what it costs at each tier, and how it stacks up against Cursor, Copilot, Aider, Devin, and the rest of the 2026 agentic coding field.

What Claude Code Does

A single session can:

• Read any file in your project
• Edit files via a structured diff tool (so you can review before anything is written)
• Run shell commands, scoped by permission prompts you configure
• Run tests and observe failures, then iterate on a fix
• Search the web for documentation
• Use MCP servers to talk to external systems (databases, GitHub, Slack)
• Spawn subagents for parallel work (research, code review, planning)
• Apply skills — reusable expert procedures defined by markdown files
• Commit, push, and open pull requests through the gh CLI

That list looks like a job description for a junior engineer, and that comparison is deliberate. The tool was built to replicate the loop a human developer runs constantly: read code, form a hypothesis, make a change, run a check, read the output, adjust. What makes it "agentic" rather than just "autocomplete with extra steps" is that it decides which of those actions to take next on its own, inside guardrails you set, rather than waiting for you to specify every command.

Anthropic's Opus 4.8 update (released May 28, 2026) added "dynamic workflows" to Claude Code — a mode that can orchestrate hundreds of parallel subagents inside a single session to handle codebase-scale migrations spanning hundreds of thousands of lines, from planning through merge, using the project's existing test suite as the pass/fail bar.

The model under the hood is Anthropic's most capable coding model at any given time. Through 2025 that was Claude Opus 4.5 (released November 24, 2025), which Anthropic described as the best model in the world at the time for coding, agentic work, and computer use, and which led SWE-bench Verified — a benchmark of real, unresolved GitHub issues pulled from production repositories — among frontier models available at launch. By mid-2026 the lineup had moved through Opus 4.6, Opus 4.7 (April 16, 2026), and Opus 4.8 (May 28, 2026), each shipping alongside specific Claude Code features rather than as a silent backend swap.

How the Agentic Loop Actually Works

Understanding Claude Code means understanding the difference between completion and agency. A completion tool predicts the next few tokens of code based on what's already on your screen — that's what powers inline autocomplete in most editors. An agentic tool instead runs a loop: it reasons about a goal, picks a tool call (read a file, run a test, search the web), observes the result, and decides what to do next. It keeps looping until the goal is met or it needs a decision only a human can make.

Think of the difference the way you'd think about a GPS versus a driving instructor sitting in the passenger seat. A GPS (completion) tells you the next turn based on where you are right now. A driving instructor (an agent) watches you drive, notices you're heading toward traffic, suggests an alternate route, and if you miss the turn anyway, adjusts the plan rather than just repeating the old instruction. Claude Code is built to be the instructor, not the GPS.

How does Claude Code decide what to do next in a session?

It runs an observe-plan-act loop: it reads relevant files and command output, forms a short-term plan, executes one tool call at a time (edit a file, run a command, search the web), and re-evaluates after each result. Longer or riskier actions — like force-pushing or deleting files — require explicit permission unless you've pre-approved that category of command.

Subagents: how parallel work happens

A single Claude Code session runs one primary thread of reasoning, but it can spin off subagents — isolated instances with their own context window, system prompt, and tool access — to handle sub-tasks in parallel. A subagent might be told "review this diff for security issues" or "research how this library's API changed between versions," and it reports back a compressed summary rather than dumping its entire working process into the main conversation.

This matters because context windows are a finite, shared resource. Every tool call, file read, and command output eats into the space the model has to reason about your actual problem. Delegating exploratory or noisy work to a subagent keeps the primary thread focused, the way a lead engineer delegates "go figure out why staging is down" to a teammate instead of doing the digging personally while also trying to review a pull request. For the full mechanics and comparisons to frameworks like AutoGen and CrewAI, see our AI subagents explainer.

By 2026, subagent orchestration in Claude Code had grown from a novelty into infrastructure: Anthropic's Opus 4.8-era "dynamic workflows" mode can run hundreds of subagents concurrently inside one session for large migrations. That scale creates its own cost trade-off, covered below.

Skills: reusable expertise, not just prompts

A skill in Claude Code is a markdown file that packages a specific procedure — how your team writes database migrations, how to format a changelog entry, how to run your project's particular deploy script — and gets auto-triggered when the agent judges it relevant to the current task. Skills are the difference between an agent that knows programming in general and one that knows your codebase's conventions specifically.

Hooks and CLAUDE.md: encoding project memory

Two features do most of the work of making Claude Code behave consistently on a given project:

• CLAUDE.md — a file at the project root that Claude Code reads at the start of every session. It holds standing instructions: deployment rules, naming conventions, "never touch this directory," architectural notes. Anthropic and this site's own engineering workflow both treat CLAUDE.md as the single source of truth a project's agent sessions should defer to.
• Hooks — shell commands that fire automatically on lifecycle events, such as running a formatter every time a file is edited or a linter before a commit. Hooks turn "please remember to run prettier" into something that happens without being asked.

Key Features at a Glance

Table: Claude Code's core capabilities and what each one solves

Feature	What it does	Problem it solves
Subagents	Spawn isolated agents in parallel for sub-tasks; results return as summaries	Keeps the main context window focused during long, exploratory tasks
Skills	Reusable expert prompts auto-triggered by relevance (markdown-defined)	Encodes project- or team-specific procedures the base model doesn't know
Hooks	Run shell commands on lifecycle events (e.g., format on file edit)	Enforces conventions without manual reminders
MCP servers	Plug in any MCP server for external tools	Standardizes how the agent reaches databases, GitHub, Slack, and other systems
Slash commands	User-defined commands like /review or /deploy	Turns multi-step routines into a single invocation
CLAUDE.md	Project-level instructions read every session	Gives the agent standing context instead of re-explaining rules each time
IDE integration	VS Code, JetBrains, Cursor extensions for inline use	Lets terminal-first workflows surface inside a familiar editor
Background tasks	Long-running operations with completion notifications, including GitHub Actions support	Frees the developer from babysitting slow jobs
Dynamic workflows (2026)	Orchestrates large numbers of subagents for codebase-scale work	Makes multi-hundred-thousand-line migrations tractable in one session

Pricing in 2026

Claude Code isn't a separate subscription — it's included with a Claude.ai plan or billed through the metered API, and Anthropic's official pricing documentation lists the following consumer and business tiers:

• Pro plan (roughly $17–$20 per month) — modest daily usage limits; workable for a few hours of engineering work per day or for hobbyist projects.
• Max plan ($100–$200 per month, depending on tier) — substantially higher usage limits and support for multiple parallel sessions, aimed at engineers running Claude Code most of the working day.
• Team plan (roughly $20–$125 per seat, depending on standard vs. premium tier) — adds collaborative admin controls for small organizations.
• Enterprise — SSO, audit logs, custom rate limits, and negotiated volume pricing.
• API, metered per token — useful for CI pipelines, headless automation, and scripts that don't need a human in the loop. As of Anthropic's mid-2026 pricing page, Claude Opus 5 and Opus 4.8 API access runs $5 per million input tokens and $25 per million output tokens; Claude Sonnet 5 runs $2 per million input and $10 per million output tokens through an introductory period ending August 31, 2026, after which it moves to $3/$15.

The practical shape of the decision: Pro suits a few hours of engineering per day, Max suits all-day work with several sessions running at once, and API metering suits automated pipelines where a human isn't watching the terminal. One caveat worth taking seriously: Anthropic's own subagent documentation notes that subagent-heavy workflows can add substantial token overhead — multiple parallel workers each consuming their own context — compared to running the same task as a single linear agent. That overhead is the cost of the speed-up; it's not free parallelism.

Claude Code vs Other AI Coding Tools

Table: How Claude Code compares to the major AI coding tools of 2026

Tool	Surface	Mode	Strength	Weakness
Claude Code	Terminal + IDE plug-ins	Agentic	Large-codebase navigation, long agentic loops, big refactors	Terminal learning curve, no native GUI
Cursor	IDE (VS Code fork)	Mostly autocomplete + Agent mode	Best inline editing UX	Less aggressive at open-ended, long-horizon loops
GitHub Copilot	IDE	Autocomplete + Workspace + Agents	Deep IDE and GitHub-native integration	Historically less context-aware across large repos than Claude Code or Cursor
Aider	Terminal	Constrained agentic	Lightweight, bring-your-own-model	Smaller scope, fewer built-in integrations
OpenAI Codex CLI / Agents SDK	Terminal	Agentic	Strong GPT-model integration	Younger agentic-coding ecosystem
Windsurf	IDE (Codeium fork)	Agentic	"Flows" show multi-step plans visibly	Smaller user base than Cursor or Copilot
Devin (Cognition)	Cloud + Slack	Fully autonomous cloud agent	Async, "fire and forget" background work	Less interactive; can be costlier per task
Replit Agent	Cloud IDE	Agentic	Beginner-friendly, runtime included	Tied to Replit's hosted infrastructure

The general pattern by mid-2026: Cursor still wins for tight, interactive editing inside a familiar IDE window, Claude Code wins on long-horizon agentic tasks and large-codebase reasoning, and Devin wins for genuinely unattended background work. Most professional teams run at least two of these tools side by side rather than standardizing on one.

How does Claude Code differ from Cursor specifically?

Cursor is IDE-first: it's optimized for fast inline suggestions and a chat panel next to your code, with an Agent mode layered on top. Claude Code is terminal-first and built around longer autonomous loops — read, plan, act, verify, repeat — across many files in one uninterrupted run, which suits large refactors and multi-file bug hunts better than moment-to-moment editing assistance.

Do I need an Anthropic API key to use Claude Code?

No. Claude Code works with a consumer Pro or Max subscription through Claude.ai login. An API key is only necessary for metered pay-as-you-go usage, typically in CI pipelines or scripted automation where no person is present to authenticate interactively. Other autonomous coding agents include Devin, Cognition's AI software engineer and Cline, the VS Code AI coding agent.

What Claude Code Is Good At

• Large codebase navigation. It can search, read, and reason across hundreds of files without losing track of how they relate.
• Refactors. Renames, type migrations, and framework upgrades that touch dozens of files at once.
• Agentic loops. "Find the failing test, fix it, commit, push" runs end-to-end without a human re-typing each step.
• Documentation. Reading code and producing accurate READMEs, runbooks, and changelogs from the actual implementation rather than stale notes.
• Multi-tool workflows. Combining git, gh, package managers, Docker, and MCP servers inside a single task.
• Large-scale migrations. With Opus 4.8's dynamic workflows, orchestrating parallel subagents across a codebase-scale change and using the existing test suite as the quality bar.

What Claude Code Is Less Good At

• No graphical interface. The terminal is comfortable for engineers but a real barrier for non-developers or teams that live in visual tools.
• Pixel-level visual design. Like other large language models, it struggles to judge fine visual layout without being shown screenshots to react to.
• Real-time multi-user collaboration. Sessions are single-user by design; two people can't co-drive one session the way they might in a shared document.
• Cost discipline on heavy subagent use. Parallel subagent workflows can consume meaningfully more tokens than the equivalent single-threaded task, so API users especially should set budgets rather than assume more parallelism is automatically cheaper or faster overall.
• Judgment on ambiguous product decisions. It will make a reasonable technical call when given a vague instruction, but "reasonable" and "what your team actually wanted" aren't guaranteed to be the same thing — review still matters.

Why Claude Code Took Off

Two forces converged. First, Anthropic's coding models were near or at the top of SWE-bench Verified — a benchmark built from real, previously unresolved GitHub issues rather than synthetic coding puzzles — through 2025 and into 2026, meaning the underlying model was genuinely strong at production-style engineering problems, not just toy examples. Anthropic's Opus 4.5 announcement, for instance, reported it led SWE-bench Verified among frontier models at its November 2025 release and led SWE-bench Multilingual across seven of eight languages tested.

Second, Anthropic shipped the agentic harness — subagents, skills, hooks, MCP, background tasks — faster than most competitors turned equivalent raw model capability into a usable workflow. A powerful model without the scaffolding to let it read files, run commands, and iterate independently is just a very good autocomplete. The harness is what converts model quality into hours of engineering time actually saved, and Anthropic's own testing showed early sessions completing in one pass what had previously taken 45 minutes or more of manual work.

For broader context on the concepts underneath Claude Code, see our pieces on AI agents and AI subagents.

Common Misconceptions About Claude Code

"It writes production code with no review needed." It doesn't, and Anthropic doesn't market it that way. Claude Code's structured diff tool exists specifically so a human reviews every proposed change before it's applied — treating agent output like any other pull request, not like a black box you merge blindly, is the workflow Anthropic recommends and the one serious engineering teams actually use.

"More subagents always means faster results." Parallelism helps when sub-tasks are genuinely independent, but each subagent carries its own context overhead, and coordinating many of them adds real token cost. For a small task, a single linear agent is often both cheaper and faster than spinning up a swarm.

"It only works for greenfield projects." In practice, its stronger use case is often the opposite: legacy codebases with tests, where "read the existing pattern and extend it consistently" plays to an agent's strength at reading large amounts of code quickly, more than it plays to a human's patience for the same task.

"You need to be a command-line expert to use it." Basic terminal literacy helps, but the tool is designed to be conversational — you type what you want in plain English, and the agent translates that into the actual commands. The learning curve is real but shallower than "expert-level shell scripting."

FAQs

Q: When was Claude Code released? A: Anthropic introduced Claude Code as a limited research preview in February 2025, alongside the Claude 3.7 Sonnet announcement, and later moved it to general availability with native VS Code and JetBrains integrations and GitHub Actions support for background tasks.

Q: How is Claude Code different from Cursor? A: Cursor is an IDE-first tool built around inline editing and autocomplete with an added Agent mode. Claude Code is terminal-first and biased toward long agentic loops — multi-step plans that read, edit, run, and verify across many files in one continuous session.

Q: Do I need an Anthropic API key? A: No. Claude Code works with the consumer Pro or Max subscription through your Claude.ai login. API keys are an alternative for metered, pay-as-you-go usage — typically in CI pipelines or headless scripting.

Q: What are subagents, and why do they matter? A: Subagents are specialized agents the main Claude Code session can spawn to handle isolated tasks in parallel, each with its own context window and tool access, reporting back a summary. See our AI subagents explainer for the full architecture and how it compares to frameworks like AutoGen and CrewAI.

Q: Can Claude Code talk to my database or GitHub directly? A: Yes, through MCP servers — Model Context Protocol, Anthropic's open standard for connecting models to external tools. Anthropic and the community maintain reference MCP servers for Postgres, GitHub, Slack, the filesystem, and many other systems.

Q: Is Claude Code suitable for production codebases? A: Yes, with the same review discipline you'd apply to any contributor. Use a CLAUDE.md file to encode project conventions, scope permissions rather than granting blanket shell access, and treat every agent-authored change as a pull request that needs a human review before merge.

Q: How much does Claude Code cost? A: It's bundled with a Claude.ai Pro (roughly $17–$20 per month) or Max ($100–$200 per month) subscription, or billed per token through the Anthropic API for automated or CI use. There's no separate Claude Code-only price tag.

Q: What underlying model powers Claude Code? A: Whichever Claude model Anthropic considers its strongest coding option at the time — Claude Opus 4.5 through late 2025, then Opus 4.6, Opus 4.7, and Opus 4.8 across 2026, each release paired with specific Claude Code feature updates rather than a silent swap.

Q: Can Claude Code run tasks in the background without me watching? A: Yes. Background tasks, including GitHub Actions integration, let long-running operations continue while you work elsewhere, with a notification when the task completes.

Q: What's the biggest risk of using Claude Code carelessly? A: Skipping review. The tool is built with a diff-based edit flow and permission prompts specifically so a human stays in the loop; bypassing that by granting broad, unscoped shell permissions or merging changes unread reintroduces the exact risk the design is meant to prevent.

The takeaway

Claude Code succeeded by pairing a genuinely strong coding model with an agentic harness — subagents, skills, hooks, MCP, background tasks — that turns raw model capability into hours actually saved on real engineering work. It isn't a replacement for review discipline, and heavy parallel subagent use has a real cost, but for large-codebase navigation, refactors, and long multi-step loops, it does something autocomplete-style tools were never built to do: work independently toward a goal and check its own progress along the way.

---

Read next:

• What Is Vibe Coding? AI-First Programming
• What Is MCP? Model Context Protocol Explained
• AI Subagents Explained — Parallel AI Work

Sources

• Anthropic, "Claude 3.7 Sonnet and Claude Code" (February 2025), anthropic.com/news/claude-3-7-sonnet
• Anthropic, "Introducing Claude Opus 4.5" (November 24, 2025), anthropic.com/news/claude-opus-4-5
• Anthropic, "Introducing Claude Opus 4.7" (April 16, 2026), anthropic.com/news/claude-opus-4-7
• Anthropic, "Introducing Claude Opus 4.8" (May 28, 2026), anthropic.com/news/claude-opus-4-8
• Anthropic, "How Claude Code is used in practice," anthropic.com/research/claude-code-expertise
• Anthropic, Claude Platform pricing documentation, platform.claude.com/docs/en/about-claude/pricing
• Anthropic, Claude Platform release notes, docs.anthropic.com/en/release-notes/overview
• SWE-bench Verified leaderboard, princeton-nlp/SWE-bench