Claude Code

Claude Code is Anthropic's coding agent. It reads your repository, edits files, runs commands and talks to your tools, from a terminal, VS Code, JetBrains, a desktop app or the web. It needs a paid Claude plan, a Console account or a cloud provider; plans start at $17 a month on annual billing. Facts checked 31 October 2026.

At a glance

What Claude Code is, in plain terms

Most people meet a coding agent as a chat box in an editor. Claude Code started somewhere else: in the terminal, where you type a request, it plans the work, touches the files it needs and runs your tests. Anthropic's overview says it "understands your entire codebase and can work across multiple files and tools to get things done."

It has since spread to other places. The docs say each surface connects to the same engine, so the CLAUDE.md files, settings and MCP servers you set up once work in all of them. That matters if you switch between a terminal at your desk and a browser tab on a train.

Typical jobs the docs list:

• writing tests for untested code, then running them and fixing failures
• fixing lint errors across a project, resolving merge conflicts, updating dependencies
• building a feature from a plain-language description
• creating commits, branches and pull requests
• reviewing pull requests in CI

Here is a one-shot example straight from the overview:

claude "write tests for the auth module, run them, and fix any failures"

Where it runs

You can move a session between them. claude --teleport pulls a web session into your terminal (needs a claude.ai subscription), and /desktop hands the current terminal session to the desktop app on macOS and x64 Windows.

Install and first run

The native installer is the recommended route.

# macOS, Linux, WSL
curl -fsSL https://claude.ai/install.sh | bash

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

# Homebrew (stable channel)
brew install --cask claude-code

# WinGet
winget install Anthropic.ClaudeCode

Three details from the docs are worth knowing before you pick one:

1. Native installs update themselves in the background. Homebrew and WinGet installs do not; you run brew upgrade claude-code or winget upgrade Anthropic.ClaudeCode yourself.
2. Homebrew has two casks. claude-code follows the stable channel, "typically about a week behind", and skips releases with major regressions. claude-code@latest gets new versions as soon as they ship.
3. On native Windows, Git for Windows is recommended so Claude Code can use the Bash tool. Without it, PowerShell is the shell tool.

After installing, open a new terminal and check it worked:

claude --version
cd your-project
claude

On first use it asks you to log in. If ANTHROPIC_API_KEY is set and you approve the key, it skips the login prompt. The docs also list apt, dnf and apk packages for Debian, Fedora, RHEL and Alpine.

Plans and prices

Claude Code is not sold on its own. It comes with Claude plans, with API billing, or through a cloud provider. The consumer plan table on the pricing page, checked 31 October 2026:

Prices exclude tax, and Anthropic says plans are "subject to change at Anthropic's discretion". The pricing page also has Team and Enterprise tabs and an API tab. For a longer breakdown see Claude pricing and the Max plan page.

Two practical points. Limits reset on a rolling five-hour window and a weekly cap, and how fast you hit them "depends on the model you use, how long your chats are, and how large your files are." A big agentic session on the top model eats usage faster than a short edit on a small one. And API billing is a separate path: you pay per token with no seat price, which suits CI jobs and bursty use better than heavy daily use.

Models you can use

Claude Code runs Claude models. The model-configuration docs give these aliases:

The alias can point at an older version on other providers. On Amazon Bedrock and Google Cloud's agent platform the docs show sonnet resolving to Sonnet 4.5, for example. Effort levels (low through max) are also configurable, and the docs list which models support which levels.

Pro lists Fable as "paid with usage credits", while Max includes it up to half your weekly limit. If you care about the model name on a given plan, read the pricing page again on the day you subscribe; this part changes often.

MCP, skills, hooks and subagents

This is where Claude Code is deepest, and where most of the "best add-ons" questions come from.

MCP. The Model Context Protocol is the open standard for linking an agent to outside tools. Adding a remote server is one line:

claude mcp add --transport http notion https://mcp.notion.com/mcp

The docs also cover local stdio servers, .mcp.json project files and directory connectors. Our Claude Code MCP servers list ranks the ones people install first.

Skills. Skills package repeatable workflows your team can share, such as /review-pr or /deploy-staging. See skills.

Hooks. Shell commands that run before or after Claude Code actions, like formatting after every file edit.

Subagents. Several agents working on different parts of a task, with a lead agent that assigns work and merges results. Background agents let you run full sessions in parallel and watch them from one screen. See subagents.

Agent SDK. If you want your own agent built on the same tools, the docs point to the Agent SDK; our Claude Agent SDK page covers it.

Add-ons worth knowing first

Automation beyond the terminal

The overview lists several ways to run Claude Code without sitting at it:

• Pipes and scripts. claude -p runs a single prompt non-interactively, so you can feed it logs or a diff: git diff main --name-only | claude -p "review these changed files for security issues".
• CI. GitHub Actions and GitLab CI/CD for PR review and issue triage.
• Scheduled tasks. Cloud scheduled tasks (formerly routines) run even when your computer is off and can fire on API calls or GitHub events. Desktop scheduled tasks run on your machine. /loop repeats a prompt inside a CLI session.
• Chat. Mention @Claude in Slack with a bug report and get a pull request back.
• Browser. A Chrome integration for debugging live web apps.

Limits, privacy and data handling

The data-usage page is blunt about what depends on your account type.

Some other facts from that page:

• Prompts and model output travel over the network encrypted with TLS 1.2 or higher. Claude Code runs locally, but the model does not.
• Session transcripts are stored in plaintext on your machine under ~/.claude/projects/ for 30 days by default. The cleanupPeriodDays setting changes that.
• Transcripts you send with /feedback, /bug or /share are retained for 5 years.
• The optional transcript upload after the session survey keeps data for up to 6 months, and nothing uploads unless you choose Yes.

Other limits are practical ones. It is Claude-only, so you cannot point it at a Gemini or GPT model through a picker. It is closed source, so you cannot patch it. And because it can run shell commands, permission settings matter; the docs have a settings and permissions section that is worth reading before you give it a production machine.

Who it suits

Claude Code fits you well if:

• you work in a terminal, or you want to script an agent in CI
• you already pay for a Claude plan and want the agent included
• you want a deep customization story: skills, hooks, subagents, plugins
• you need one agent across terminal, IDE and phone

It fits less well if you want to swap between vendors' models in one screen, if you need an open-source licence, or if you want a flat price with no usage window. Those cases point to the alternatives below.

Moving to it from another agent

Switching is cheaper than it sounds, because the instruction file is the main thing to carry over.

1. Instructions. If your repo has an AGENTS.md for another agent, Claude Code can read it. Otherwise copy the content into a CLAUDE.md at the project root.
2. MCP servers. Re-add each one with claude mcp add, or move the entries into a .mcp.json file in the project. Servers that follow the standard work across clients.
3. Rules and snippets. Editor-specific rule files do not carry over. Fold the useful ones into CLAUDE.md or turn repeated ones into skills.
4. Try it inside your editor first. The VS Code extension also installs in Cursor, so you can run both in one window and compare on the same task before you commit to anything.

Alternatives and duos

Confused about the name? "Claude" is the chat product and "Claude Code" is the coding agent; see Claude vs Claude Code. For the full field, start at AI coding agents.

Our take

This section comes from the docs and pricing pages, not from a benchmark run. We have not tested Claude Code for this page.

What stands out in the documentation is breadth. One engine is exposed through a CLI, two editor families, a desktop app, the web, Slack and CI, and the instruction file follows you between them. Few agents document that many entry points with a shared configuration.

The price model is the thing to think about. A $17 or $20 plan is a low bar to start, but heavy agentic use runs against a five-hour window and a weekly cap, and the jump to $100 is large. If your use is bursty, API billing may suit better. If it is daily, the plan ladder is the realistic choice.

The trade-offs are the two named above: Claude models only, and closed source. If either is a deal-breaker, an open-source terminal agent such as OpenCode or Codex CLI is the place to look. If neither is, the practical test is to install it and run it on one real task in a repository you know well.

Checklist before you commit

• Confirm which plan includes the model you want on the pricing page today
• Decide whether your data-training setting is on or off
• Put your build, test and style rules in CLAUDE.md
• Add one MCP server you will really use, not five
• Review permission settings before using it on shared machines
• Try the same task in one alternative, so you have a comparison on your own code