What is Claude Code?

Claude Code is Anthropic's AI coding agent. It reads your codebase, plans fixes or new features, writes the actual code, and checks its own work. Unlike a chatbot that you'd have to prompt turn by turn, it runs on an agentic loop that locks an objective and keeps working until it meets it.

Claude Code vs. Claude

First, a little history (sorry). Claude Code was designed specifically for developer workflows. But when it launched, non-technical users saw the advantage: they could use an AI agent on top of a folder on their computer to execute tasks such as web research or creating documents in bulk.

This led to Anthropic releasing Claude Cowork in January 2026. It applied the same agentic principle of Claude Code to regular office work, with easy connectors to popular tools to make it interoperable.

Over time, it became frustrating for Claude Cowork users to have to decide whether a task was better for a regular chat or for an agent. So in September 2026, Anthropic decided to merge Cowork into the regular Claude experience—it now fires the agentic features when it detects that you need them.

If you're not strictly coding, the regular Claude app might be a better place to get started. It can do all the regular chatbotty tasks, plus it can take care of more complex tasks like:

Organizing and cleaning up files. Point it to a folder and sort files out, remove duplicates, and convert file formats in bulk.
Working on spreadsheets and data. Move data in and out of spreadsheets, analyze it, and build reports or dashboards on top of it.
Drafting documents and slide decks. Add your company data and easily get documents and slide decks that always follow the same formatting and aesthetic conventions with Claude Docs and Claude Slides.

That being said, if you're an advanced non-technical user, you might find Claude Code to be worth the trouble. It does better than regular Claude when you need to bridge different platforms without native connectors or for difficult jobs that require handling multiple files or heavier automation.

Whether you go with regular Claude or Claude Code, you can extend Claude's abilities with Zapier, which lets you securely access and act on 9,000+ apps directly from the chat window. Claude even recommends Zapier as a solution when it doesn't have a native connector for the use case.

Try Zapier

What can you do with Claude Code?

The million-token question. Claude Code helps you code simple tools, put together websites, build apps, configure web services, and everything in between, just using prompts. Many professional developers use it every day at work to build new features or fix bugs, while semi-technical operators pick it to assemble scripts, tools, and prototypes.

Here are some examples of Claude Code in the wild.

Going from idea to prototype in an afternoon. Swiss bank Pictet struggled with innovation: central teams were building with AI, while everyone else was only interacting with chatbots. The bank organized a hackathon to source ideas for solving an employee access rights problem; teams started working with Claude Code and submitting their proposals with real logic in each prototype. The best submission was a step ahead of previous attempts and ended up being implemented as the main tool.
Accelerating coding and testing. Japanese company Classmethod struggled with not having enough hours for strategic development work. The team decided to integrate Claude Code into their dev lifecycle, automating code generation from GitHub issues, with self-review included to catch security risks and code quality issues. This led to an 80% decrease in time spent reviewing code.
Rebuilding a user interface quickly. HubSpot needed to refresh its frontend for the 2025 rebrand. The engineering team put Claude Code to work to refactor the frontend code to fit the new brand direction, collapsing months of manual work.
Non-technical people building tools and shipping code. Atlassian saw the productivity gains from Claude Code in its engineering teams, and wanted to push that to non-technical teams. In a pilot program, product managers and designers used the coding agent to prototype, build custom internal apps, and commit small code changes. The Zapier team also had a building challenge for the entire marketing team, and many of the pilots are now running in production.

How to install Claude Code

Desktop app

The Claude desktop app gives you access to both the regular chatbot experience as well as Claude Code. It's the best for interacting with the agent using a graphical interface natively on your device. If you're non-technical, this is the way to go.

Visit the Claude website to download the latest version.
Once you finish installing it, open the app and log in to your Claude account.
On the top-left of the main menu, click the Code icon (usually represented as </>).
Use the controls around the prompt input to choose your repo.

That's it! You're ready to build.

Web

Claude Code is available as a web app, which is helpful for working across multiple devices, managing cloud routines, and working on cloud repos.

Head to claude.ai, and log in to your account.
Click the Code icon in the top left. It should look like this: </>.
Pick a cloud repo to begin (you can't work on local repos using the web app).

Start chatting with it, and you're good to go. Everything runs on the web, and code outputs on the web are saved in the cloud. Claude creates PRs for the cloud repo that you connect, so make sure to merge the changes as you build.

Tip: For automating tasks in the cloud, try Routines: you can run them on a schedule or based on triggers, connect external data sources, and share code by connecting a GitHub repo.

IDE extensions

Claude Code also offers IDE extensions for VS Code and all JetBrains IDEs. It offers an experience close to pair programming, minus code autocomplete.

In VS Code, you get a side-tab chat to prompt and manage Claude Code. It automatically uses your open file as context, and you can mention files in your folder with the @ sign or select lines of code to send to the model. It can also show diffs side-by-side or within the chat stream. Get started with Claude Code VS Code extension by searching for it in the extensions panel or via the VS Code marketplace.
In JetBrains, you get a quick-launch shortcut that opens a terminal version of Claude Code—it's not an integrated tab in the interface. It can also share context from your project, supports file references, and displays diffs in the native viewer, but it's not as powerful as its VS Code sibling. Install the Claude Code extension in any JetBrains IDE by heading to Settings > Plugins or via the JetBrains marketplace.

Terminal

The terminal is the original way to run Claude Code. It's also the most flexible—and the absolute coolest.

On macOS, Linux, or WSL, open your terminal and run the following: curl -fsSL https://claude.ai/install.sh | bash. (On Windows, open PowerShell and run the following instead: irm https://claude.ai/install.ps1 | iex.)
Confirm it installed with claude --version.
cd into your project folder and run claude to start your first session. You'll be prompted to log in to your Claude account the first time.

You're ready to prompt.

The terminal is also where headless mode lives: add the -p flag to any command (like claude -p "fix the failing test"), and Claude runs that one prompt without UI or an interactive back-and-forth. That's what makes it possible to drop Claude Code into scripts or cron jobs, and it's the building block for setting up loops that run unattended in the background.

How to use Claude Code

Setup

You'll start by pointing Claude Code to either a local folder or a cloud repository. After you've locked one in, the next best step is to set up a CLAUDE.md file at the root of the folder and fill it with the instructions and conventions for the project you're working on. It works like a system prompt, and it'll save you from repeating yourself every time you start a new thread.

If you're not sure how to do either of these things, just ask Claude Code, and it'll guide you through the process.

Prompting

After you set up CLAUDE.md, you can send in your first prompt: ask for anything from a product requirement document for an idea you had to a prototype of a web app. Depending on what you need, you can start in Plan mode to flesh out an approach before executing, or rely on Auto mode to start building right away.

As soon as it receives your request, Claude starts thinking about how to approach it.

Usually, it checks what's already in the folder by reading file names, sometimes reading the full file contents if it believes it matches the task. For unclear prompts, Claude can stop to ask questions about design and functionality. They'll show up as multiple-choice questions.

Multiple-choice questions for the user in Claude Code

After you send your answer, the agent continues moving forward. As it does, you'll notice it's using multiple tools such as:

Bash for navigating and interacting with your file system
WebSearch and WebFetch to browse the web and read web pages
Create/Read/Write/Edit to work on and manipulate files

Unlike a chatbot, each tool in the agent harness—the infrastructure around the AI model that allows it to stay on task and take actions—allows the model to use your computer, run code, and interact with external services.

MCP servers

Speaking of external integrations, Claude Code includes an MCP client, which means it can connect with MCP servers to use their capabilities or pull resources. Once you log in to the external tool, the agent can take any approved actions. For example, it can configure Supabase databases from scratch or update your documentation in Notion. If Claude Code doesn't have a native connector for the app you need, Zapier MCP securely extends it to 9,000+ apps.

Keep in mind that every input and output of using these tools is stored in the context window, along with the thinking stream. You can keep adding more prompts to control how Claude is building your solution, but keep an eye on the context usage: a fuller context window increases usage and makes the model slowly lose focus on the task, increasing the likelihood of hallucinations or mistakes. You can easily use /clear to clear the context window.

Skills

At some point, you'll find yourself doing repeatable tasks, such as turning bug reports into requirements or testing your code. This is where agent skills come in: you can create a skill to automate tasks, and call it with a slash command whenever you need it (or whenever the model detects that intent).

Those skills are stored locally in the .claude/skills/ folder in your repo; if you want them available across all your projects, you can save them to ~/.claude/skills/.

Subagents and loops

Some tasks benefit from parallelization: having two or more agents exploring a solution or running two independent but related tasks at the same time.

You can call subagents just by stating it in any prompt: how many you want and what they should do. The main agent spins them up, briefs them on their respective tasks, and then aggregates the results when the subagents finish their tasks.

Each subagent has its own context window and only returns its final answer to the main agent. This is good when you don't want to poison the main agent's context window with irrelevant or conflicting information, but it's not ideal when you want to access the context window of the subagent. Use this awareness to choose when to use two main agent threads in parallel (you keep context on both) versus one main agent managing multiple subagents (context for subagents is discarded).

Subagents can talk to one another to collaborate on complex tasks. For example, you can have one subagent setting up a database while another brainstorms a data schema; in this case, when the database configuration agent finishes, it can pass the access information to the schema builder so it can complete its setup.

But some advanced users don't even spend time looking at the interface at all. They spend their time designing loops, where you can set an agent to keep working by itself to hit a specific goal, using all the resources it has available. It grabs requirements, fleshes them out, turns them into new code or edits to existing code, and tests it before logging a PR. Human engineers then spot-check critical parts of the code, using automated CI/CD tools to assess code quality or structural problems.

What's new in Claude Code?

Anthropic ships Claude Code updates almost every week. Here are some of the major ones we've seen over the past year:

Auto mode skips "are you sure?" prompts. The new Auto mode skips asking for permission on nearly every single action the agent takes. It's now the default, but you can still switch to Manual, Plan, or Accept Edits if you prefer. The overall building experience feels faster; Claude only stops when it encounters an important issue.
One screen for managing all agents. Run claude agents to know what's running, what's waiting on you, and what's done. Subagents run in the background, while Claude keeps working on your main task. Sessions can message each other, sharing key points of context to keep work moving forward.
Set a goal and walk away. Type /goal followed by a finish line, like "all the tests pass." Claude works turn after turn without waiting for your next prompt, while a fast model checks if your goal was reached. The process stops if the goal is met, if a model finds it impossible to complete, or something breaks that only you can fix (for example, an empty Anthropic credit balance).
Scheduled cloud tasks. Routines, available on Claude Code web, run a saved prompt on a schedule or based on a trigger, using a cloud repository as a base for the action.
Larger context windows. The latest Sonnet, Opus, and Fable models all come with a 1M-token context window.
Better analytics for skills and usage. If you've built a few skills, /skill-doctor shows how often you used each and how much it costs in context tokens. Typing /usage breaks down your current usage and also what actions are devouring your plan's limits.

Claude Code pricing

Claude Code comes with your Claude subscription, so you can jump right in if you have one. You can also skip the subscription and pay per token through the API.

Plan	Price	What you get for Claude Code
Free	$0	Not included
Pro	$20/month billed monthly, or $17/month billed annually ($200 upfront)	Access to Opus and Sonnet
Max	Starts at $100/month	Choose 5x or 20x the usage of Pro, plus higher output limits and priority access at busy times
API	Pay as you go	No usage caps: - Sonnet 5.5: $2/M input tokens and $10/M output tokens - Opus 5.5: $4/M input tokens and $20/M output tokens

Claude Code shares its usage limits with the rest of your plan, and those limits reset on a rolling five-hour window. That means a long afternoon in Claude Code and a day of regular chatting draw from the same pool. On top of the five-hour window, there are weekly caps that reset on a fixed day each week, so you can't just burn through everything in one sprint.

You'll feel the ceiling in long sessions, running several subagents in parallel, or big /goal loops. If you're hitting your limit every day, that's the signal to upgrade to a more expensive paid plan. The API makes sense for unpredictable usage or running Claude Code in scripts; but since this is pay-as-you-go, keep an eye on your billing dashboard.

Claude Code best practices

Claude Code comes with a few limitations that you'll want to keep in mind before you dig in. The good news is that there are solutions (sometimes in the form of workarounds), so here are some best practices.

Use /clear to free up context. I'm not alone in having usage limit issues; online discussions are constantly bringing it up, with users reporting burning through their allowance quickly. The 1M-token context window is part of the issue: /clear your session after you finish a task to free up context—otherwise the entire history keeps burning tokens.
Always review outputs. Despite Claude's autonomy while coding and checking its own work, you still need to review the outputs to make sure it didn't make mistakes. Keep an eye on it especially as it works in Auto mode, as sometimes it can drift from the task or get stuck in overengineering a solution.
Learn the nuances before you start prompting. Don't let the ease of use fool you: there's a learning curve to Claude Code. And you'll find that vague prompts get expensive because the model will assume (and build) everything that's not explicitly stated. So I recommend learning how to set up your CLAUDE.md well, understanding how to use Plan mode, and being careful about when it's finally time to pull the trigger and let the agent build on its own.

Take action from Claude Code with Zapier

Claude Code can connect to your other apps through MCP, but it gets messy fast. Every app needs its own server, its own sign-in, and its own permissions. And some apps don't have a decent MCP server yet.

Zapier MCP replaces the mountain of connectors with a single pipeline. It exposes Claude Code to 9,000+ apps, including favorites like Slack and Linear and thousands of niche apps across industries. As soon as you connect your work apps to Zapier, you can ask Claude to run actions on them.

How to connect Claude Code to Zapier MCP

In your terminal, run claude mcp add --transport http "Zapier-MCP" https://mcp.zapier.com/api/v1/connect.
Sign in with your Zapier account when prompted. Zapier creates and configures the MCP server during sign-in, so you don't need to set one up first.
Restart Claude Code, and ask it to run the list-skills tool. Find the onboarding skill, and follow it to switch on the apps you want Claude to use.
Type /mcp to confirm the connection. You should see Zapier listed as connected, along with how many tools it exposes.

Prefer a plugin? Run /plugin marketplace add zapier/zapier-mcp first (if you haven't already), then run /plugin install zapier@zapier-plugins.

You decide what Claude can touch, and you pick which apps and actions are available: read-only lookups, sending messages, or updating records. Sign-in happens through OAuth, so you don't have to manage secrets. Plus, it runs on your Zapier tasks, so there's no separate MCP bill. Each successful tool call uses two tasks from your plan, and failed calls are free.

If you need something that runs on a schedule, with branching logic, loops, and error handling, the Zapier SDK is a better match. Claude Code can write code that calls Zapier's apps directly. The authentication layer is already figured out for you, and the SDK generates types that tell the AI coding agent what each action expects as inputs, making it faster to reach the solution.

Try Zapier

Is Claude Code for you?

If you write code for a living, yes: Claude Code is one of the most popular AI coding agents right now. It's great for big multi-file changes, scaffolding new projects, and bug fixes that touch half a codebase. See how it stacks up versus Cursor and Codex, if you're still shopping around.

If you don't code, it depends on the job. For sorting files, interacting with spreadsheets, or drafting documents, the regular Claude app gets you there with less setup. But if you're willing to climb a steeper learning curve to build your own scripts and small internal tools, Claude Code gives you more control than regular Claude. If you already pay for a Claude subscription, give it a try and see if you vibe.

Claude Code FAQ

Is Claude Code free?

Claude Code isn't available on Claude's free plan; you'll need at least a Pro subscription to access it. The Free tier gives you Claude chat on web and mobile, but the terminal access and agentic features that make Claude Code useful are gated to paid plans. If you want to test it before committing, one month of Pro at $20 is the cheapest way in.

How is Claude Code security?

Unless you're still wearing your AI-shaped tin foil hat (in which case, why are you reading this article?), Claude Code has you covered for security. Anthropic encrypts data at rest (AES-256) and in transit (TLS 1.2+), and holds SOC 2 Type II and ISO 27001 certifications. For enterprise customers, there's a zero-data-retention option. On the local side, Claude Code uses strict read-only permissions by default and asks for your explicit approval before it edits files, runs tests, or executes commands. The main thing to keep in mind: just like with any other agent harness, anything you point Claude Code at ends up in the context window, so be deliberate about which folders and files you expose to it.

How much does Claude Code cost?

Claude Code is included in all paid Claude plans. Pro is $20/month (or $17/month billed annually) and covers most individual users who aren't running it for hours on end every day. If you're hitting the usage ceiling regularly, because of long agentic loops, parallel subagents, or big codebases, Max 5x ($100/month) or Max 20x ($200/month) gives you proportionally more capacity per five-hour session. If you'd rather pay as you go without a subscription, the API charges per token. That can be useful for scripts or unpredictable usage, but watch your billing dashboard.

Related reading:

This article was originally published in August 2025 by Maddy Osman. The most recent update was in September 2026.

What is Claude Code?

Everything you need to know about Claude Code—including my personal review.