Claude Code is a coding tool built by Anthropic that works agentically: instead of just suggesting the next line of code, it reads a project’s files, plans a multi-step approach, edits code, runs shell commands and tests, and reports back — asking for approval before it changes anything, by default. It runs in a terminal, but also as an extension inside code editors, a standalone desktop app, and a browser-based version.

How it’s different from a chatbot or autocomplete

Claude, the assistant covered in our explainer on Claude, answers questions in a chat window. Claude Code is a different product built on the same underlying Claude models, aimed at a different job: instead of copying an answer out of a chat, a developer points it at a real codebase and lets it act inside it directly — opening files, making edits, running a test suite, and using git. That is also what separates it from inline-autocomplete tools such as GitHub Copilot in their original form: Claude Code is built around multi-step, semi-autonomous task completion rather than suggesting a single next line, though the wider category of AI coding assistants increasingly blurs that line as competitors add their own agent modes.

How it works

Under the hood, Claude Code runs an “agentic loop”: it reads the relevant files, decides on a plan, calls a small set of built-in tools — reading and editing files, running terminal commands, searching the codebase, using git — and checks its own output, repeating until the task is done or it needs a decision from the developer. For larger jobs, it can also split work across several sub-agents that each handle one part of a task and report back to a lead agent coordinating the whole.

Two features extend that basic loop. A CLAUDE.md file at the root of a project is read at the start of every session, letting a team record coding standards, build commands, and architecture notes once instead of repeating them in every request. And the Model Context Protocol, an open standard Anthropic introduced, lets Claude Code connect to outside systems — ticket trackers, design documents, databases, or a company’s internal tools — so it isn’t limited to what’s already on disk.

For anything that touches files or runs a command, Claude Code shows the proposed change and asks for permission before applying it, unless a developer explicitly switches on an “accept all” mode for a session — the guardrail that lets it be trusted with real production code.

Why it matters

The practical effect of an agentic coding tool is scale: a review that would take a person days or weeks of manual work can be handed to an agent that works through a codebase file by file, at machine speed. That is the property behind Alberta’s government using Claude Code to scan 466 million lines of code for vulnerabilities in 20 hours — a task it estimated would otherwise take 6.5 years of manual review. The same scale advantage shows up in less dramatic, everyday work: clearing a backlog of lint errors, writing missing tests, resolving merge conflicts, or updating a dependency across hundreds of files in one pass.

How to start

Claude Code installs from a terminal command (curl -fsSL https://claude.ai/install.sh | bash on macOS, Linux, and WSL, or a PowerShell/WinGet equivalent on Windows), and also ships as a VS Code and JetBrains extension, a desktop app, and a browser version at claude.ai/code. Anthropic’s own quickstart guide walks through installing it and running a first task. It requires a Claude account: it’s included with a Claude Pro subscription — $17/month billed annually or $20/month billed monthly, as of July 2026, per Claude’s pricing page — as well as the higher-usage Max plans, or it can be used through the Anthropic Console at standard per-token API rates for teams that prefer usage-based billing over a seat price.

FAQ

Is Claude Code free?
There’s no standalone free tier. It’s included with Claude’s paid Pro, Max, Team, and Enterprise plans, or usable through the Anthropic Console at API rates.

Does it need an internet connection?
Yes — it sends requests to Anthropic’s models to plan and generate each step, even though the files it reads and edits stay on the local machine (or a repo it’s pointed at).

Can it use models other than Claude?
The terminal CLI and VS Code extension support some third-party model providers, but Claude Code is built around Anthropic’s own Claude models by default.

Is it safe to let it edit real code?
It asks for approval before changing files or running commands unless a developer turns on an auto-accept mode, and it works alongside normal git workflows, so any change can still be reviewed or reverted like a human’s.