Claude Code: a terminal coding agent you install with a shell script

What Claude Code does that a chat window does not

A chat window answers questions about code you paste into it. Claude Code is described in the README as an agentic coding tool that lives in the terminal, understands your codebase, and executes routine tasks, explains complex code, and handles git workflows through natural language commands. The distinction is where it runs and what it can touch. It operates from your project directory, so the repository is the context rather than a pasted excerpt. The README also lists two other surfaces: an IDE, and tagging @claude on GitHub.

The intended user is someone who already has a checkout open and a terminal running. If your work happens in a browser tab or a design tool, the terminal-first model adds a step rather than removing one. If your work is mostly reading unfamiliar code and committing small changes, the fit is closer, because those are the tasks the README names.

How the agent is structured: CLI, plugins and hooks

The repository layout shows more than a single binary. Alongside the README there are directories named plugins, examples, mods and scripts, plus a .claude-plugin directory and a .claude directory at the top level. The README states that the repository includes several Claude Code plugins that extend functionality with custom commands and agents, and points to plugins/README.md for the available ones. That means the shipped behaviour is partly composed from plugin definitions rather than being entirely compiled into the tool.

The examples directory contains subdirectories named gateway, hooks, mdm and settings. Those names map onto configuration surfaces: hooks for event-driven behaviour, settings for project or machine configuration, mdm for managed device deployment, and gateway for a network path. The README itself does not explain any of these four, so treat the directory names as a map of what exists, not as documentation. For anything beyond installation and the plugins note, the README defers to the official documentation at code.claude.com/docs/en/overview.

A changelog file sits at the top level, and the recent releases are numbered in the v2.1.x range, with v2.1.251 published on 2026-08-28. The last push to the default branch was on 2026-08-28 as well.

Installing Claude Code without the deprecated npm path

The README puts a note above the install list: installation via npm is deprecated. The recommended methods come first. On macOS and Linux the documented one-liner downloads and runs an install script.

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

Running this pipes a remote script straight into bash, so if your environment requires inspecting scripts before execution, download it first and read it. Homebrew is listed as the other macOS and Linux option.

brew install --cask claude-code

On Windows the README gives a PowerShell equivalent and a WinGet package.

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

winget install Anthropic.ClaudeCode

The npm route still exists but is labelled deprecated, so it is the one to avoid unless you have a specific reason.

npm install -g @anthropic-ai/claude-code

After installing, the README's second step is the whole workflow: navigate to your project directory and run the command. There is no project scaffolding step and no config file to author before the first run.

claude

You should see the tool start in the context of that directory. The README does not document what the first-run prompt looks like, so expect to follow the setup documentation for authentication details. For uninstall steps and troubleshooting, the README points at code.claude.com/docs/en/setup rather than describing them inline.

Where Claude Code is the wrong tool

The README states that usage data is collected, including code acceptance or rejection, associated conversation data, and feedback submitted through the /bug command. It also states that safeguards include limited retention periods for sensitive information, restricted access to user session data, and policies against using feedback for model training. Those are policy statements with links to the data usage page, the Commercial Terms of Service and the Privacy Policy. If your code cannot leave your machine under any policy, this design is a mismatch, and no amount of configuration described in the README changes that.

The second boundary is the licence. The repository has a LICENSE.md at the top level, but the licence identifier is not given in the metadata available here. Anyone evaluating this for a company should open that file rather than assume. A repository that is not archived and had its last push on 2026-08-28 tells you the code is moving; it tells you nothing about what you are permitted to do with it.

The third boundary is scope. The README frames the tool around executing routine tasks, explaining code, and git workflows. It does not claim to replace a test runner, a linter, or a CI system. A team looking for deterministic, reproducible automation should keep that automation and treat the agent as an interface on top of it, not a substitute.

How it differs from editor-integrated assistants like Copilot

GitHub Copilot is the closest comparison most engineers will reach for, and the difference is where the loop closes. Copilot-style assistants complete code inside the editor buffer and leave running commands, reading files across the tree, and committing to you. Claude Code is described as living in the terminal and handling git workflows itself, which moves the boundary from suggestion to action.

That shift has a cost. An inline completion is easy to ignore. An agent that edits files and runs git commands needs review before you accept its output, and the README's feedback description implies exactly that kind of accept-or-reject loop. The repository also shows an IDE surface and a GitHub mention surface, so the comparison is not strictly terminal versus editor; it is agent-with-repository-access versus completion-in-buffer.

If your team's bottleneck is typing speed inside a file, the completion model is the smaller change. If the bottleneck is the sequence of steps around the code, such as finding the right files, explaining an unfamiliar module, or assembling a commit, the agent model addresses the part that completion does not.

Support, bug reports and what the repository does not document

Bug reports have a documented path: the /bug command inside Claude Code, or a GitHub issue on the anthropics/claude-code repository. The README also points to a Claude Developers Discord for help and feedback. A SECURITY.md file exists at the top level for security reporting, though the README does not describe its contents.

What is missing from the README is as informative as what is present. There is no documented rollback procedure for a version upgrade, no statement about how plugin updates are versioned, and no explanation of the gateway, hooks, mdm or settings examples beyond their directory names. The upgrade cost is therefore hard to estimate from this repository alone: releases arrive frequently within the v2.1.x line, and the README does not say whether the install script pins a version or tracks the latest. Teams that need to freeze a version should confirm that behaviour against the setup documentation before standardising on the installer.

On licensing, the repository carries a LICENSE.md, and the README links to the Commercial Terms of Service and Privacy Policy for data handling. Read those directly rather than relying on any summary, including this one.

Editorial conclusion

Adopt Claude Code if you already work in a terminal or an IDE and want an agent that can read the repository and run routine git and editing tasks from natural language. Skip it if you need an offline tool, a self-hosted model, or a licence you can read in the repository before installing. Before rolling it out to a team, verify two things first: the setup documentation at code.claude.com/docs/en/setup, which the README names as the place for uninstall steps and troubleshooting, and the data usage page at code.claude.com/docs/en/data-usage, because the README states that usage data, conversation data and /bug feedback are collected. The npm install path is marked deprecated in the README, so pick a non-npm method unless you have a reason to stay on it.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes