Claude Code: Anthropic's Agentic Coding Tool, Explained

Most coding assistants wait for you to paste a snippet and ask a question. Claude Code does the opposite: you describe the outcome in plain language, and it reads the codebase, makes a plan, edits the files, runs the tests, and opens the pull request. It lives in your terminal, it has passed 153,000 stars on GitHub, and it is now the tool Anthropic's own engineers use to build Claude itself. That last point is not marketing gloss — Anthropic's S-1 filing, published in October 2026, says Claude wrote more than 90% of the code contributions to the company's codebase in September 2026, and Claude Code is the tool that did much of that work.

None of that makes it free, simple, or safe to run blind. There is no free tier, no standalone price, and the usage limits it runs on are shared with everything else you do in Claude. This is the full picture: what Claude Code does, how to install it, what each way of paying actually costs, and where the catches are.

Quick answer

Claude Code is Anthropic's agentic coding tool for the terminal, IDE, desktop app and web. It understands your codebase, plans multi-step work, edits files, runs commands and opens pull requests, all from natural-language instructions. It is included in every paid Claude plan — Pro at $20 a month, Max from $100 — or through a Console account billed per token. It is not free, it is not open source, and its extension system runs with your full user privileges, so what you install matters as much as what you type.

What Claude Code actually is

Anthropic's own description is short: an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows through natural-language commands. The word doing the work there is "agentic". A chatbot answers; an agent loops. Claude Code gathers context, takes action, and verifies the result — then repeats until the task is done or it needs you to decide something. That loop is the same engine behind every surface it runs on: terminal, VS Code, JetBrains IDEs, the Claude desktop app, and the web at claude.ai/code.

The GitHub numbers show how fast that idea caught on. The repository, anthropics/claude-code, was created on February 22, 2025, and by October 28, 2026 it had 153,642 stars and 26,063 forks, with the latest release, v2.1.312, shipping on October 27. Releases land almost daily — eight shipped between October 7 and 16 alone — which is both a strength and a warning: the tool improves fast, and the documentation you read last month may already be out of date.

One common misunderstanding deserves clearing up first. Claude Code's code is visible on GitHub, but it is not open-source software. The repository carries no licence file, and GitHub's own API reports no licence for it. You can read the code; the repository does not grant the usual open-source freedoms. That distinction matters if your company has rules about what "open source" means in a dependency review — and it is a contrast with the community harnesses built on top of Claude Code, which are usually MIT-licensed.

What it can do

The everyday jobs are the ones the product page leads with: explain an unfamiliar codebase, fix a bug, turn a GitHub issue into a pull request, and run a refactor that takes hours. The workflow Anthropic demonstrates is a bug fix, not a code snippet — Claude reproduces the fault, traces it to the root cause, fixes it, and explains what changed. Plan mode is the other half of that: for larger work, Claude writes out a plan first and waits for your approval before touching anything.

Beyond the terminal, the surfaces share one engine, so your CLAUDE.md files, settings and MCP servers carry across all of them:

• Terminal: the full CLI. Run claude in a project directory and start typing. Headless mode (claude -p) runs it from scripts and CI.
• VS Code and JetBrains: extensions with inline diffs, @-mentions, plan review and conversation history. Cursor is supported through the same extension search.
• Desktop app: the Code tab in the Claude desktop app, with multiple sessions side by side and scheduled tasks. A paid subscription is required.
• Web: cloud sessions at claude.ai/code, for long-running tasks and repos you do not have locally, available on desktop browsers and the Claude mobile apps.
• GitHub and GitLab: tag @claude on an issue or pull request, or run it in GitHub Actions and GitLab CI/CD for automated review and triage.
• Chat and messaging: a Slack integration routes bug reports to pull requests, and Channels push events from Telegram, Discord, iMessage or your own webhooks into a session.
• Remote Control and claude --cloud: start a task locally, continue it from your phone, and pick it up on another machine.

Model choice is a command, not a setting buried in a menu. /model opens an interactive picker; claude --model opus starts a session on a specific model. Sonnet 5.5 is the default — the everyday model for implementation, tests and refactoring. Opus 5.5 is for the hardest work: architecture, thorny debugging, long-running agents. Haiku 5.5 is the fast, cheap option for exploration, searches and small edits. There is also an opusplan alias that plans with Opus and executes with Sonnet, so the expensive model only does the thinking.

How to install it

Installation takes one command, and Anthropic has a clear preference for which one. The npm install (npm install -g @anthropic-ai/claude-code) is deprecated; the native installer is recommended.

On macOS or Linux:

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

Or with Homebrew:

brew install --cask claude-code

On Windows, in PowerShell:

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

Or with WinGet:

winget install Anthropic.ClaudeCode

Then move into a project and run claude. The first run asks you to sign in with a Claude Pro, Max, Team or Enterprise plan, or with a Console account. If you have set the ANTHROPIC_API_KEY environment variable and approve it when asked, the login prompt is skipped. The docs note two practical details: the installer shows no progress while it downloads, and on Windows, Git for Windows is recommended so Claude Code can use Bash instead of PowerShell as its shell tool. Native installs update themselves in the background; otherwise claude update does it.

Two free resources are worth knowing about before you pay for a course anywhere else. Claude Academy runs free self-paced courses including Claude Code 101, and the docs ship a quickstart that walks through a first real task, from exploring a codebase to committing a fix.

Skills, subagents, hooks, MCP and plugins

The extension system is where Claude Code stops being a single tool and becomes a platform — and where the security story gets serious. There are five layers, and they stack:

• Skills are reusable instructions stored in a project's .claude/skills folder or your user directory. They load only when invoked, so they cost no context until they are used. Your project's memory lives in CLAUDE.md files, which Claude reads automatically; if a project has no CLAUDE.md, an AGENTS.md file is honoured too.
• Subagents run in their own context windows, so a research agent or a reviewer does not pollute your main conversation. Two sessions on the same machine can even message each other.
• Hooks fire on 33 events — before a tool runs, after it finishes, when a prompt is submitted, when a model switches — and can block a tool call, rewrite its arguments or replace its result. Since v2.1.295, a hook can set onFailure: "block", so a security hook that crashes or times out stops the action instead of waving it through.
• MCP servers connect Claude Code to outside tools — GitHub, databases, monitoring — through the Model Context Protocol, configured per project in .mcp.json.
• Plugins bundle skills, subagents, hooks and MCP servers into one installable package. Anthropic's official plugins live in a separate claude-plugins-official repository.

In October 2026 a sixth layer arrived: mods. A mod is a plugin written in JavaScript or TypeScript that runs inside Claude Code itself, rather than as an external script. A mod can draw a pane next to the transcript, hold a tool call while it asks a question, or send a request to another model — and some of Anthropic's own features, including the /diff panel, now ship as mods.

Here is the part to read twice. A plugin or mod runs with your user privileges: it reads your secrets, sees every prompt, and can approve a tool call on your behalf. Anthropic's documentation states plainly that mods are not sandboxed. The same applies to skills that use hooks — security researchers showed in October 2026 that a malicious skill could execute outside the normal permission protections. Install extensions only from sources you trust, pin the versions you run, and treat a random skill from the internet the way you would treat a random shell script from the internet — because that is effectively what it is.

What it costs

Claude Code has no standalone price. You pay for it through one of three routes, and they bill very differently. Prices below were read from claude.com/pricing and Anthropic's API pricing page on October 28, 2026. Prices exclude tax, and Anthropic notes they can change.

Subscriptions: Pro and Max

The fine print that catches people out is the shared pool. Usage limits reset on a rolling five-hour session window, with weekly limits on top, and the pool is shared across Claude on the web, desktop, mobile and Claude Code. A long chat in the browser eats the same allowance as a coding session. There is no fixed message count: how much you can do depends on the length and complexity of your conversations and the models you choose. When you hit a limit you wait for the reset, move to a higher plan, or switch on usage credits, which let you keep working past the plan limit at standard API rates. Usage credits are managed with /usage-credits, and you can set a monthly spend limit on them — do that before your first heavy week, not after it.

Console: pay per token

With a Claude Console account, Claude Code consumes API tokens at standard rates, with no monthly ceiling unless you set one. The current model line-up and rates:

All four current models have a 1M-token context window. Cache reads cost 5% of the input price on Sonnet and Opus, which is why long sessions stay affordable — the repeated context is nearly free. Batch API runs cost half the standard rate. And if you are on a data-residency plan, responses are billed at 1.1x the list price, which Claude Code folds into the session cost figure.

What does a real month look like? Anthropic's own cost guidance puts the enterprise average at around $13 per developer per active day and $150–250 per developer per month, with costs below $30 per active day for 90% of users. That is an average, not a promise: a developer running several Opus agents in parallel will blow past it, and a developer doing short Sonnet sessions will sit well under it. The tooling to find out where you land is built in — /usage shows session cost, cache statistics and which skills, subagents and MCP servers your tokens went to; /insights writes an HTML report on how you actually work; and --max-budget-usd caps a single session's spend. If you want the wider field before you commit, our roundup of the 7 best AI coding agents in 2026 compares Claude Code's pricing and limits against Codex, Cursor, Copilot and the rest.

Pros and cons

What stands out: it is genuinely agentic — it plans, acts and verifies, rather than autocompleting your next line. It runs locally and talks directly to Anthropic's APIs, with no backend server and no remote code index, which is a cleaner privacy story than tools that upload your repo for indexing. The permission model asks before it edits files or runs commands. The cost tooling — /usage, /insights and budget caps — gives you per-skill and per-server token attribution, which is more visibility than most rivals offer. And it is developed in the open on GitHub at a pace — near-daily releases — that most competitors do not match.

What to watch: there is no free tier and no way to try it without a paid plan or a Console account. The shared usage pool means your chat and your coding compete for the same allowance. API billing has no ceiling unless you set one. Extensions run unsandboxed with your privileges, so a bad install is a real security event, not a nuisance. And the repository is source-visible, not open source — do not treat it as one in a compliance review.

FAQ

Is Claude Code free?

No. The Free plan does not include it. The simplest route is Pro at $20 a month, or $17 a month if you pay annually ($200 up front). The other route is a Console account, where you pay per token and there is no monthly fee — but also no ceiling unless you set one.

Can Claude Code use models other than Claude?

On a Claude subscription or Console account, it runs Anthropic's models. The Terminal CLI, VS Code and JetBrains surfaces also support third-party providers, per Anthropic's own docs — but that is a bring-your-own-key setup, not a supported Anthropic offering, and the experience varies by provider.

Is Claude Code open source?

No. The code is visible on GitHub under anthropics/claude-code, but the repository has no licence file, so it does not carry open-source freedoms. Community projects built on top of it, such as plugin packs and workflow harnesses, are usually licensed separately — check each one.

Does it work in my IDE?

Yes. There are extensions for VS Code and JetBrains IDEs, support for Cursor through the same extension search, a Code tab in the Claude desktop app, and a web version at claude.ai/code for cloud sessions. Every surface shares the same engine, CLAUDE.md files and MCP servers.

How much will heavy use actually cost?

Anthropic's published enterprise average is about $13 per developer per active day, or $150–250 a month, with 90% of users under $30 a day. Treat that as a starting estimate: run a two-week pilot, watch /usage, set a usage-credit spend limit, and use --max-budget-usd on unattended sessions.

What data does Anthropic collect from Claude Code?

The README states Anthropic collects feedback including usage data — such as code acceptance or rejection — and associated conversation data. It also states safeguards: limited retention periods for sensitive information, restricted access to session data, and a policy against using feedback for model training. Paid plans show model training as opt-out on the pricing page.

Key takeaways

• Claude Code is a terminal-first agent that plans, edits, tests and opens pull requests — not a chatbot with a code window.
• Entry costs $20 a month for Pro ($17 on annual billing); Max starts at $100; or pay per token through a Console account.
• Sonnet 5.5 is the default model; Opus 5.5 handles the hardest work and Haiku 5.5 handles the cheap, fast jobs.
• Usage limits are shared across all of Claude, so a long chat and a long coding session draw from the same pool.
• Skills, plugins and mods run unsandboxed with your privileges — install them only from sources you trust.

Claude Code is a serious tool for developers who live in the terminal, and the price of entry — a Pro subscription — is lower than the cost of the time it saves on a single real task. The honest way to evaluate it is a two-week pilot on Pro: watch /usage, keep the expensive models for the work that earns them, and decide with your own numbers rather than anyone else's benchmark. As of October 28, 2026, that is a $20 experiment with a clear answer at the end of it.