Claude Code vs Claude.ai: which Claude are you actually using?

What it actually is

Anthropic ships Claude in two main forms, and the names are close enough to trip anyone. Claude.ai is the chat: the website plus the desktop and mobile apps, where you type or paste into a box and read the reply. It writes code happily, and it can build small interactive demos (artifacts) right inside the conversation. But everything stays in the chat window. If you want that code in your project, you copy it out yourself.

Claude Code is an agent. You point it at your project folder, and it reads your files, edits them, creates new ones, and runs commands like npm install or your test suite, asking your permission as it goes. That is agentic coding in practice: the AI works in your real environment instead of handing you snippets to paste.

Think of it as texting a plumber versus having one over. Texting a photo of the leak (claude.ai) gets you a diagnosis and instructions you carry out yourself. The house call (Claude Code) is someone opening the cabinet and touching the actual pipes while you approve each step. The expertise on the other end is the same, because both products run the same Claude models. What changes is where the work happens.

	Claude.ai	Claude Code
Shape	Chat app (web, desktop, mobile)	Coding agent (terminal, IDE, desktop app, browser)
Your files	Sees only what you paste or upload	Reads and edits your project directly
Runs commands	No	Yes, with your permission
Best for	Thinking, learning, quick prototypes	Building and fixing a real project

Why your AI just did this

If you asked claude.ai to “fix the bug in my app” and it answered that it cannot access files on your computer, you hit the boundary between the two. The chat knows only what you paste. So it asks for one file, then another, then the error message again, and twenty minutes in you have become a human file system. Claude Code removes that loop. It reads the repository itself, follows the trail across files, and checks its own work by running the code.

The confusion flows the other way too. A tutorial says “have Claude build it,” you open claude.ai, and the instructions suddenly mention a terminal you were never asked to open. Or your agent refers to CLAUDE.md, the instruction file Claude Code reads at the start of every session, which the chat never sees. And tools like Cursor can run Claude models inside their own agent, which is a third thing again: the model inside is Claude, the agent around it belongs to Cursor.

When you’ll run into it

The classic first sighting is typing claude in your terminal and getting command not found: claude. Paying for Claude does not install Claude Code. It is a separate install (curl -fsSL https://claude.ai/install.sh | bash on Mac and Linux, with Windows installers too), and you log in with the same account you use for the chat.

You will also meet the name in more places than you expect. Claude Code runs in the terminal, inside VS Code and Cursor through an extension, in JetBrains IDEs, as a desktop app, and in the browser at claude.ai/code. “Claude in a browser tab” no longer tells you which product you are in.

Then there is the bill. On Pro and Max plans, one subscription covers both products, and everything you do in either one counts against the same shared usage limits. A long afternoon of chatting on claude.ai leaves less room for an evening Claude Code session. How those limits are measured at all is covered in tokens vs credits.

What to check

• Check which window you are in before you paste: a chat box on claude.ai touches nothing on your machine, while a Claude Code prompt acts on your real files.
• Log in to Claude Code with the same account you use on claude.ai, so your Pro or Max plan covers it instead of a separate payment method.
• Run /usage inside Claude Code, or open Settings > Usage on claude.ai, to see how much of the shared pool is left before you start something big.
• Look for a stray ANTHROPIC_API_KEY environment variable if surprise API charges show up: when that variable is set, Claude Code bills the key per token instead of drawing on your subscription.
• Split the work on purpose: think through ideas and rough plans in the chat, then bring the decided version to Claude Code, where the agent can make the edits and verify them.

Say it like a dev

Instead of: “I pay for Claude, why do I need to install something else?” Say: “My Pro plan covers both the claude.ai chat and Claude Code, but Claude Code is a separate install. I’ll set it up and log in with the same account.”

Instead of: “Claude can’t see my files” Say: “The claude.ai chat only sees what I paste. I’ll open this project in Claude Code so the agent can read the repo and make the edits itself.”

Instead of: “which Claude is smarter?” Say: “They run the same models. The chat is for talking a problem through, the agent is for changing the code.”