How to use Claude: web, desktop, mobile, terminal and API

Claude is the AI you can chat with in a browser, talk to on your phone, hand a chore to in Chrome — and, if you write software, let loose on your codebase from the terminal. It is one of the most capable AI systems in the world, and one of the most interesting to learn, because it shows up in so many places.

This guide covers all of them. We start with who makes Claude and how it is trained, choose between its models, and then walk through each way of using it step by step: the web, desktop and mobile apps, Projects and Artifacts, Claude in Chrome, Claude Code in the terminal, the Model Context Protocol, and finally the Claude API in code.

What is Claude?

Claude is a family of AI models made by Anthropic, and the assistant you talk to that runs on them. You write or speak in ordinary language; Claude writes back. It reads documents and images, searches the web, builds charts and small apps, works through spreadsheets, and writes and runs code.

Under the hood Claude is a large language model: a neural network trained on a huge amount of text to predict what comes next, then trained further to be a helpful, honest assistant. (Our first article, Welcome to the world of AI, explains that process from scratch.)

Anthropic’s story

Anthropic was founded in 2021 by a group of former OpenAI researchers who wanted to build frontier AI with safety research at the centre. Claude is its product and, in a sense, its argument: that an AI can be both very capable and trustworthy.

How Claude is shaped: Constitutional AI

Most assistants are tuned with human feedback: people rate answers, and the model learns to give the kind they prefer. Anthropic added a twist it calls Constitutional AI. Claude is also guided by a written set of principles — a “constitution” — about being helpful, honest and avoiding harm. During training the model critiques and revises its own answers against those principles, and AI feedback based on them is used alongside human feedback.

You will notice the results in everyday use. Claude tends to tell you when it is unsure, points out problems with a plan instead of flattering it, explains why it will not help with something harmful, and pushes back politely when you are wrong.

The model family

Claude comes in sizes. Anthropic’s long-standing naming runs from small to large, like poetry: Haiku (short and fast), Sonnet (balanced) and Opus (a major work). The Claude 5 family adds Fable for the most demanding work. In the apps you pick a model from a menu; in the API you name it in code.

All current Claude models read images as well as text. The larger ones can hold about a million tokens in their context window — hundreds of thousands of words, enough for several books or a sizeable codebase in one conversation. Newer models also think adaptively: they decide for themselves how long to reason before answering, and developers can steer that with an effort setting.

Every way to use Claude

Which door you use depends on what you want to do. Here is the whole map at a glance, then an interactive helper.

Step by step: Claude on the web

1. Open claude.ai and create a free account with your email or a Google account.
2. Tell Claude about yourself. In Settings, add a line or two about who you are and how you like answers — for example, “I’m a teacher; keep things practical and use Indian examples.” Claude takes this into account in every chat.
3. Learn the screen. It has only a handful of parts:

4. Ask for something real, then keep the conversation going: “shorter”, “more formal”, “give me three options”, “now as a table”. Each reply builds on everything before.
5. Drop in files. Drag a PDF, spreadsheet, photo or screenshot into the message box and ask about it.
6. Use Research (on paid plans) when you need a proper answer from many sources: Claude searches for several minutes and returns a report with citations.

Projects and Artifacts

Artifacts: Claude builds things beside the chat

Ask for a document, a diagram, a chart or a small interactive tool and Claude creates an Artifact — a finished piece of work shown in a panel next to the conversation. You can keep chatting to change it, copy it, download it or publish it to share.

Projects: a workspace for one subject

A Project bundles files and instructions so you do not have to repeat them. Upload your syllabus, company handbook or research papers once, add a few lines of instructions, and every chat in that Project can draw on them.

Claude on desktop and mobile

Desktop. Download the Claude app for Mac or Windows from claude.com/download. It has everything the website has, plus a keyboard shortcut to summon Claude from anywhere, access to local connectors on your computer, and a built-in way to use Claude Code without a terminal.

Mobile. Install the official Claude app by Anthropic from the App Store or Google Play and sign in with the same account; your chats sync everywhere. The phone is where two features shine: the camera (photograph a notice, a plant, a maths problem, a menu in another language) and voice mode, a spoken conversation that is great for language practice, interview rehearsal, or thinking aloud on a walk.

Claude in Chrome

Claude in Chrome is a browser extension that puts Claude in a side panel. It can see the page you are on, and — with your permission — click, type and navigate to get things done: compare products across tabs, fill in a long form, pull figures out of a dashboard, or tidy a to-do list in a web app.

Prompting Claude well

Everything from the six ingredients in ChatGPT 101 applies: role, task, context, format, examples and limits. Three habits work especially well with Claude:

1. Be explicit and give the reason. Current Claude models follow instructions closely. “Keep it under 100 words — it’s for a phone notification” beats “be brief”, because Claude can use the reason to make better choices.
2. Structure long prompts with XML-style tags. Wrapping pasted material in labelled tags such as <document> and <task> makes clear what is material and what is instruction.
3. Put long documents first and your question last. With big inputs, leading with the material and ending with the ask tends to give better answers.

You are reviewing a rental agreement for a first-time tenant in Bengaluru.

<agreement>
  …paste the full agreement here…
</agreement>

<my_situation>
  Moving in 1 Nov. Plan to stay 2 years. Have a cat.
</my_situation>

<task>
  List clauses that are unusual or unfavourable to me,
  quoting each one. Then give 3 questions to ask the owner.
</task>

Claude Code: Claude in your terminal

Claude Code is Anthropic’s agentic coding tool. Instead of pasting code into a chat, you run Claude inside your project. It reads your files, searches the codebase, edits code, runs commands and tests, and works through multi-step tasks — asking your permission before anything risky. It runs in the terminal, in VS Code and JetBrains IDEs, in the desktop app and on the web.

Install and start

Claude Code needs a Pro, Max, Team or Enterprise plan, or an API (Console) account. On macOS, Linux or WSL:

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

# or with Homebrew
brew install --cask claude-code

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

cd ~/projects/my-app
claude          # first run opens your browser to log in

Then just describe what you want. A typical session looks like this:

$ claude
✳ Welcome to Claude Code

> The checkout page shows the wrong total when a coupon is applied. Find and fix it, and add a test.

● Read src/cart/total.ts
● Search "applyCoupon" — 3 matches
● Read src/cart/coupon.ts

The discount is applied before GST in total.ts but after it in coupon.ts,
so it is counted twice. I’ll apply it once, before tax.

● Edit src/cart/total.ts
-   const net = subtotal - discount(subtotal) - discount(withTax);
+   const net = subtotal - discount(subtotal);

? Run npm test? › Yes   Yes, and don’t ask again   No

● Bash npm test — 42 passed (1 new: “coupon applied once”)

Fixed. The total now matches the invoice in all 3 coupon cases.

How it works

Claude Code is an agent: the model runs in a loop with tools for reading, searching, editing and running commands. Each turn it looks at the results of its last action and decides the next one, until the job is done or it needs you.

Commands worth knowing

Teach it your project

The single most useful thing you can do is keep a good CLAUDE.md: a short file of how to build, test and style this codebase, and the traps to avoid. Claude reads it at the start of every session. Beyond that, a few files let you shape Claude Code for your team:

MCP and connectors

Claude becomes far more useful when it can reach your tools and data. The Model Context Protocol (MCP), an open standard Anthropic introduced in November 2024 and now supported widely across the industry, is a common plug for exactly that. Someone writes an MCP server for a tool once — GitHub, Google Drive, a database, Slack — and any MCP-capable app can use it.

In the Claude apps these appear as connectors you switch on in settings — Google Drive, Gmail, Google Calendar, Microsoft 365 and many more. In Claude Code you add servers with claude mcp add or a .mcp.json file. Only connect servers you trust: an MCP server can act with whatever access you give it.

The Claude API

To build Claude into your own product, use the Claude API. Everything goes through one endpoint, POST /v1/messages: you send a list of messages and optional system instructions, and get Claude’s reply. Create an API key at platform.claude.com, set it as ANTHROPIC_API_KEY, and install the SDK with pip install anthropic (official SDKs also exist for TypeScript, Java, Go, Ruby, C# and PHP).

import anthropic

client = anthropic.Anthropic()  # reads your ANTHROPIC_API_KEY

message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=16000,
    system="You are a friendly science tutor for Class 8 students in India.",
    messages=[
        {"role": "user", "content": "Why is the sky blue? Suggest a simple home experiment."}
    ],
)

# A reply can hold several blocks (thinking, then text) — print the text.
for block in message.content:
    if block.type == "text":
        print(block.text)

The same request with curl, using jq to pull out the text:

curl https://api.anthropic.com/v1/messages \
  -H "content-type: application/json" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-opus-5-5",
    "max_tokens": 16000,
    "messages": [{"role": "user", "content": "Give me 3 names for a tea shop in Assam."}]
  }' | jq -r '.content[] | select(.type == "text") | .text'

Guaranteed JSON with structured outputs

When your code needs data rather than prose, set output_config.format to a JSON Schema. Claude’s reply is then guaranteed to parse and match it.

import json

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=16000,
    messages=[{
        "role": "user",
        "content": "Review: 'Room was spotless but the Wi-Fi kept dropping.'",
    }],
    output_config={
        "format": {
            "type": "json_schema",
            "schema": {
                "type": "object",
                "properties": {
                    "sentiment": {"type": "string", "enum": ["positive", "negative", "mixed"]},
                    "issues": {"type": "array", "items": {"type": "string"}},
                },
                "required": ["sentiment", "issues"],
                "additionalProperties": False,
            },
        }
    },
)

text = next(b.text for b in response.content if b.type == "text")
print(json.loads(text))  # {'sentiment': 'mixed', 'issues': ['Wi-Fi drops']}

Save money and time with prompt caching

Many apps send the same long instructions or documents on every request. Prompt caching stores that unchanging beginning so later requests reuse it: cache reads cost a small fraction of normal input and come back faster.

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=16000,
    system=[{
        "type": "text",
        "text": HR_HANDBOOK,                     # the same long text every time
        "cache_control": {"type": "ephemeral"},  # cache everything up to here
    }],
    messages=[{"role": "user", "content": "How many days of paternity leave do I get?"}],
)
print(response.usage.cache_read_input_tokens)   # > 0 on later calls = cache hit

Where to go next

• Tool use: describe your functions with JSON Schema; Claude decides when to call them and you run them and return the results — the same loop Claude Code uses.
• Built-in server tools: web search, web fetch and code execution run on Anthropic’s side, with no loop to write.
• Effort: output_config.effort trades depth of thinking for speed and cost, from low to max.
• Batches: send large offline jobs at half the price.
• Your cloud: the same models are available through Amazon Bedrock, Google Cloud and Microsoft Foundry.

Privacy and good habits

• Check important facts. Claude is careful, but it can still be wrong. Use search or Research for anything current, and open the sources that matter.
• Mind what you share. Don’t paste passwords, card numbers or ID numbers. Review the privacy settings, including whether your chats may be used to improve models, and use incognito chats for one-off sensitive questions.
• Review memory and connectors from time to time, and disconnect what you no longer use.
• At work, follow your organisation’s policy. Team and Enterprise plans have their own data terms and admin controls.
• Let it teach, not replace. For study, ask Claude to quiz you and explain your mistakes; you will learn far more than from a finished answer.

Your 30-day path

Four weeks of small steps, from first chat to first line of code.

Quick quiz

Tap a question to check your answer.

That is Claude from the first chat to the first API call. Pick the door that matches what you need today, try one real task, and build from there. ✳️

Claude and Claude Code are products of Anthropic; this article is independent and not affiliated with or endorsed by Anthropic. Model names, IDs and API prices follow Anthropic’s models overview, and install commands follow the Claude Code documentation, both as of 24 September 2026. Interface screens are illustrations and sample outputs are examples. Spotted something out of date? Tell us.