What is Claude Code?
By Mari Luukkainen · 8 min read · Last updated
On this page
- What does Claude Code do?
- What does a session look like?
- How do permission modes work?
- What is CLAUDE.md?
- What can you add to Claude Code?
- Where does Claude Code run?
- Where to go next: Claude Code questions and the page that answers each
- How is it different from Cursor, Copilot and Codex?
- What are its limits?
- Who is Claude Code for?
- Frequently asked questions
Claude Code is a coding agent from Anthropic that works inside your project folder. You describe a task in plain language, and it reads your files, edits them, runs commands such as tests and reports back. It runs in a terminal, editors, the web and a desktop app, and a permission mode sets how often it asks.
This page covers what it does, what a session looks like, how permission modes and CLAUDE.md work, where it runs and what it needs from your Claude account. For installing and daily use, continue with how to use Claude Code.
Ask AI about this pageChatGPTClaudePerplexityGoogle AI ModeGrok
What does Claude Code do?
A chat assistant gives you code in a message, and you paste it into your files. Claude Code has access to the folder you started it in, so it does the whole sequence itself:
- Look at the project: list files, open the ones that matter, search for a function name.
- Decide what to change and, if you ask, write the plan before it edits.
- Edit files, create new ones.
- Run commands: the tests, the build, a linter, Git.
- Read the output, fix what failed and run again.
- Report what changed.
Steps 4 and 5 are the difference from a chat window. The tool checks its own work against something real, a failing test or a build error, instead of handing you code that may not run.
What does a session look like?
You open a terminal in your project, type claude and press Enter. A prompt appears. You write a task:
The contact form on /contact accepts an empty message. Reproduce this with a test first, then fix it so a message under 10 characters shows the error "Write a message of at least 10 characters." Run the tests and the build. List the files you changed.The terminal shows the steps. It reads app/contact/page.tsx and the form component, proposes a test file, runs the test and shows the failure, edits the form, runs the test again and shows it pass. Last, it summarises the change. Depending on the permission mode, it may stop to ask before it edits a file or runs a command.
Writing the test first is a habit worth copying. A failing test states what "fixed" means, and the tool can check itself against it.
How do permission modes work?
Claude Code has permission modes, and the mode decides how often it asks. In manual mode it asks before each file edit and command. In auto mode a classifier approves or blocks actions for you. In plan mode it proposes and does not edit. Press Shift+Tab to cycle modes, or set the starting mode with permissions.defaultMode in settings. Per Anthropic's documentation, auto mode, where it is available, is the default starting mode in interactive terminal and VS Code sessions, and manual mode is the default for claude -p (permission modes). Check which mode your session is in before you walk away from it.
Read each command before you approve it in manual mode. Be careful with anything that deletes files, installs packages, touches Git history or sends data out. Approving a file edit with "allow for the session" lasts until the session ends. Choosing "Yes, don't ask again" for a shell command is different: it saves a rule to .claude/settings.local.json in that repository, and the command is allowed from then on, not only for the session (permissions).
A safer habit for beginners: work in a folder tracked by Git, commit before each task and review the diff at the end. Claude Code also keeps checkpoints before each prompt, and /rewind restores earlier file state. Changes made only by shell commands are not tracked by checkpoints, so Git is still the safety net.
What is CLAUDE.md?
CLAUDE.md is a text file in your project root that Claude Code reads at the start of a session. It holds what the tool should know every time: how to run the tests, where things are, your rules. Claude Code treats it as guidance, not as enforcement.
# Project: Recipe Scaler
Next.js with TypeScript and Tailwind CSS. No database.
## Commands
- Run the app: npm run dev
- Tests: npm test
- Build check before finishing any task: npm run build
## Rules
- Do not add dependencies without asking.
- Do not edit existing tests to make them pass.
- Only change files needed for the task.I use the same idea at a larger scale. I keep a private repository of markdown files, one per topic (servers, deploys, databases, monitoring, onboarding a new app), and when I open a session and ask for something non-trivial, the session reads the relevant file first. It then builds the way the rest of my setup is built, not in a slightly different way I have to reconcile later (post).
What can you add to Claude Code?
- Subagents. Separate Claude instances with their own context for a focused job. Example: a reviewer that reads a diff and reports problems, while the main session keeps working.
- Hooks. Commands that run automatically at set points, such as after Claude edits a file. Example: run the linter or type check after every edit.
- Skills. Packaged instructions Claude loads when a task needs them. Example: a checklist for publishing a release.
- MCP servers. Connections to outside tools through the Model Context Protocol. Example: give Claude read access to a database or an issue tracker.
The best practices page covers when each is worth the set-up. Each has its own guide: subagents, hooks, skills. So does plan mode.
Where does Claude Code run?
Anthropic lists these places (overview):
- The terminal CLI on macOS, Linux and Windows.
- Extensions for VS Code (also Cursor) and a plugin for JetBrains.
- A desktop app (macOS and Windows, Linux in beta).
- The web at claude.ai/code and the Claude mobile app. These start cloud sessions that run in an Anthropic-managed virtual machine, push branches and can open pull requests (cloud sessions).
- Slack, where mentioning @Claude starts a cloud session (Slack).
- GitHub Actions, using the anthropics/claude-code-action, where you mention @claude in an issue or pull request (GitHub Actions).
It needs a paid Claude plan (Pro, Max, Ultra, Team or Enterprise) or a Console account with API credits. The free claude.ai plan does not include it (setup). Plan prices, usage limits and API costs are in Claude Code pricing. Installation steps are in how to use Claude Code.
Where to go next: Claude Code questions and the page that answers each
| Question | Page |
|---|---|
| How do I install it and run a first session? | How to use Claude Code |
| What does it cost, and is there a free option? | Claude Code pricing |
| How do I keep sessions short and changes safe? | Claude Code best practices |
| What is a skill and how do I write one? | Claude Code skills |
| How do I run a command automatically on an event? | Claude Code hooks |
| How do I hand a task to a separate agent? | Claude Code subagents |
| How do I get a plan before it edits anything? | Claude Code plan mode |
| Which other tools do the same job? | Claude Code alternatives |
How is it different from Cursor, Copilot and Codex?
| Tool | Where you work | Main habit |
|---|---|---|
| Claude Code | Terminal, VS Code, JetBrains, desktop app, web | Delegate a task, review the diff |
| Cursor | Its own editor, plus a command line tool and cloud agents | Watch and accept changes as they appear, or delegate to an agent |
| GitHub Copilot | Your editor, GitHub and a command line tool | Suggestions while typing, chat, agent features |
| Codex | Command line, IDE extension, ChatGPT apps and cloud | Delegate a task, review a result |
Detailed comparisons: Cursor vs Claude Code, Codex vs Claude Code, Copilot vs Claude Code.
What are its limits?
- It can be wrong with confidence. A passing test only proves what the test checks.
- Long sessions drift. After many messages it may forget an early rule. Start a new session for a new task.
- It can run commands, so a careless approval can do real damage. Keep work under Git.
- It does not know your intent. A vague task gets a guess. See vibe coding prompts for task shapes that work.
- Security review is separate. Read is vibe coding bad.
Who is Claude Code for?
People who are comfortable with a terminal and Git, developers who want to delegate whole tasks, and non-developers willing to learn those two things. If you have never used a terminal, a browser builder is an easier first step.
I started in an editor, Cursor, in March 2025. After more than 100 apps I barely read the code anymore and the publish, build and fix loop runs fast. That is where I am now, not where to begin: each outage on my setup added a guardrail (post).
Frequently asked questions
Which Claude plan is cheapest for Claude Code?
Pro, at $20 a month or $17 a month billed annually (checked 1 November 2026). Max is $100 with 5x Pro usage, and Ultra is $200. You can also use a Console account with API credits. The free Claude plan has no Claude Code access.
Can Claude Code work while my computer is off?
Yes, through cloud sessions. Started from the web, the mobile app, the desktop app or claude --cloud, a task runs in an Anthropic-managed virtual machine, pushes a branch and lets you create a pull request from claude.ai/code. Cloud sessions are available on Pro, Max, Ultra and Team plans.
Can Claude Code use AWS Bedrock or Google Vertex instead of an Anthropic account?
Yes. Besides a Claude plan, Anthropic's documentation lists an API key, Amazon Bedrock and Google Vertex AI as ways to authenticate.
Does Claude Code read AGENTS.md?
Yes, when a project has no CLAUDE.md. If you want one file for several tools, write AGENTS.md and put the single line @AGENTS.md in CLAUDE.md.
If I write "never read .env" in CLAUDE.md, will Claude obey?
It is guidance, and nothing is blocked. To stop reads, add a deny rule such as "deny": ["Read(./.env)"] in the permissions settings, or use a hook.
Next guideHow to use Claude Code →