Claude Code — AI-Powered Coding Assistant in Your Terminal

What Claude Code is

Claude Code is a terminal-first AI coding assistant that interacts with your project at the file system level. You open a session, describe a task in plain language, and the tool reads relevant files, proposes and executes edits, runs tests, and reports back. The model powering the session is one of the Claude AI family — Opus, Sonnet, or Haiku — selected based on your account plan and any flag you pass at session start.

The defining characteristic that separates claude code from a chat window or a browser-based coding tool is that it operates directly on disk. Changes appear in your editor immediately. Shell commands run in your actual environment. Test output is real. You are not working in a sandbox or a simulated project; you are working in your repository, with your toolchain, and the AI is a participant in that process rather than a spectator suggesting code you copy elsewhere.

What Claude Code does during a session

A single claude code session can span many operations without a page reload or context reset. The sequence for a typical refactor task looks like this: the tool reads the files you describe, builds a mental model of the code structure, proposes a change plan, executes the edits, runs the affected tests, and — if a test fails — iterates on the fix until the suite passes or surfaces an error that needs human input. That cycle takes seconds per iteration rather than minutes, and the feedback loop is tight enough to feel collaborative rather than automated.

Beyond file edits, claude code can execute arbitrary shell commands: running a build pipeline, calling a linter, generating migrations, or invoking external CLIs that your project already depends on. This makes it useful for tasks that are not purely about writing code — release preparation, changelog generation, dependency audits, and documentation maintenance all sit inside the same conversational interface.

How Claude Code compares to IDE-based assistants

The comparison is worth spelling out because the use cases overlap without being identical. IDE plugins — the kind that add AI completions or chat panels inside Visual Studio Code or JetBrains products — work at the line or function level. They are fast, context-aware in a narrow window, and excellent for interactive writing where you want a suggestion every few keystrokes. Claude code works at a different granularity: multi-file, multi-step operations where the task is too large to describe as a completion prompt and the output is a set of coordinated edits across the project.

The two approaches also differ on automation. An IDE plugin assists a human doing the work; claude code can execute a complete task with minimal supervision if the instruction is clear. That difference matters for CI pipelines, batch refactors, and automated review passes, where there is no human at a keyboard during execution. Researchers at MIT CSAIL have published work on the performance characteristics of AI-assisted development workflows at scale, which provides useful grounding if you are evaluating the approach for a larger team.

Capability overview

The table below summarises what claude code can do, the relevant detail for each capability, and where to begin if you want to explore further.

Capability	Details	Where to start
File editing	Reads and rewrites source files across any language; diffs shown before commit	Features reference
Shell execution	Runs build commands, tests, linters, and arbitrary shell scripts in your environment	Shell section
Skills	Reusable capability packs (markdown + tool manifest) that extend the CLI on demand	Skills reference
Sub-agents	The CLI can spawn specialised agents for parallel or delegated work within a session	Agent docs
MCP integration	Model Context Protocol lets external tools register context providers and action handlers	MCP section
Team workflows	Shared skills, shared config, and review patterns across a squad	Teams page

Getting started with Claude Code

The shortest path to a working session is three steps: install Node.js if you do not already have a recent version, run the package install command for claude code, and authenticate with your API key. The install reference walks through the exact commands for each operating system, including the Windows PowerShell and WSL variants. Once the binary is on your path, the first session can start immediately — there is no project-level setup required, though adding a configuration file unlocks persistent preferences and team settings.

If you are evaluating claude code for a team rather than for personal use, the teams reference covers shared skill registries, per-repository config, and the review workflows that give you visibility into what the AI is changing. The enterprise page goes further with SSO, audit trails, and policy controls for regulated environments.

Model selection inside Claude Code

By default, claude code uses the model tier associated with your account. You can override this at the session level with a flag. Opus delivers the most thorough multi-file reasoning but runs at higher latency; Sonnet is the balanced choice for daily development; Haiku handles high-volume lightweight tasks at the lowest cost. The models overview keeps those comparisons current alongside context window and pricing notes.

Choosing the right model matters more than it first appears. A team running Opus for every routine task will accumulate unnecessary cost; a team defaulting to Haiku for complex refactors will hit reasoning limits that show up as incomplete or incorrect edits. The practical recommendation is to profile a typical week of work, identify which task types need long-context reasoning, and set model flags per workflow rather than per session. That one configuration decision typically cuts the weekly bill by a third without touching quality on the tasks that actually need it.