Many of us first met artificial intelligence (AI) in a chat window. We paste some code in, read the answer, and copy it back into our editor ourselves. Claude Code works another way: it opens our files, edits them, and runs our build and our tests on its own. In this lesson, we see what that means for us, where Claude Code sits among Anthropic’s products, and which other tools do the same job.

In this module you install Claude Code, sign in with your plan, and make a first small change to Repair Desk. You also learn to read the Claude Code screen and decide how much Claude Code may do without asking you.

Before we start, think back to module 0. Where is your own working copy of Repair Desk? And which two commands tell us that the solution still builds and that its tests pass?

Chat or Agent

The difference from a chat is easiest to see side by side.

An agent works in a loop. In lesson 1.4, we ask Claude Code for a new endpoint and a test for it, and it goes round like this:

1. It reads the files it needs, there Program.cs and our test file.
2. It edits both files.
3. It runs dotnet build and dotnet test.
4. It reads what they printed. If something failed, it fixes the code and goes round again.
5. When the build and the tests pass, it stops and tells us what it did.

Agentic coding is the name for this way of working: an agent makes the changes and checks them, while we set the goal and review the result. Module 2 looks closely at the loop, and at what each round costs us.

The same abilities that make an agent useful also let it do damage. An agent that edits files and runs commands can break things on our disk. It can also tell us a change works when it does not.

So we spend real time in this course on two habits.

We decide what the agent may do without asking us, starting in lesson 1.6. And we read the build and test output ourselves before we accept any change, a routine we build in module 7.

What Claude Code Is

Claude Code is the agent we use in this course, and Anthropic makes it. It is a command-line tool: we open a terminal in our project folder, start it, and type what we want in plain English. Anthropic describes it as a tool that reads your codebase, edits files, runs commands and works with your development tools. The same agent runs in several places:

• The terminal, on Windows, macOS and Linux. Every lesson in our course works there.
• Visual Studio Code and the JetBrains IDEs, through an extension or a plugin.
• The Claude desktop app, in its Code tab.
• The web, at claude.ai/code, where our session runs in the cloud.

Claude Code is not in the free Claude plan. Lesson 0.2 lists the paid plans that include it. A key for Anthropic’s application programming interface (API) works too, billed for what we use.

On every plan, a new session starts on Opus 5.5, the default model, and module 3 shows us how to choose another one.

What a Token Is

The word token comes up in almost every module of this course, so we define it once, here.

Everything Claude Code sends to the model is counted in tokens: our prompt, every file the agent reads, and every answer it writes.

Code breaks into more tokens than we might guess. In October 2026, we measured one line from Repair Desk’s Program.cs on Opus 5.5:

var today = DateTime.Now.Date;

The line has 30 characters and added about 15 tokens to our request, so code runs at about two characters per token. The whole Program.cs file, 103 lines, added about 1,500. When the agent reads a file for us, that is what the read costs.

Tokens matter to us for two reasons. First, a session can hold only so many of them: up to 1 million on Opus 5.5, Sonnet 5.5, Fable 5.1 and Haiku 5.5. Lesson 2.2 shows what fills that space.

Second, your plan’s usage limits and the prices of the API are counted in tokens.

To see the count for your own session, type /context in Claude Code. It prints how many tokens the session holds, for example 50.7k/1m tokens (5%), and which parts they come from.

Claude Code, Claude Chat and Cowork

Anthropic builds three products on the same Claude models, and we often see their names mixed up. Here is what each one is for, as of October 2026:

Two of the three are becoming one. On 16 September 2026, Anthropic started merging chat and Cowork into a single Claude app. We ask for what we need, and Claude decides whether that calls for a quick answer or a longer task.

The change reaches Pro and Max first and the other plans later, so your app may still show both. Claude Code is not part of the merge. It stays a separate tool, and it is the Code tab in that same app.

So why do we use Claude Code for this course?

Chat can explain a C# compiler error, but it cannot run dotnet test on our machine to see whether the fix works. Cowork carries out longer tasks, but it is built for documents, research and files, and it has no terminal.

Claude Code works inside our solution, with the .NET SDK (software development kit) and the git we already use. It can change our code and then prove the change with the build and the tests, and that is the skill we teach. On a paid plan, chat and Claude Code share one usage allowance.

Other Coding Agents

Claude Code has plenty of company. Most large AI companies now sell a coding agent, and they all work in the loop above. These are the main ones as of October 2026:

What differs between them is the model behind the agent, the price, the usage limits, and how each tool asks before it acts. What we teach here carries over to all of them.

We give a clear goal, plan before a big change, work in small steps, and check every result with the build and the tests.

Many of us already have GitHub Copilot. If you use it, you can keep it. Its completions and chat inside the editor do a different job from the agent we use here, and many developers use both.

Open Models and Open-Source Agents

Every agent in that table runs on its maker’s own models, in its maker’s cloud, but there is another route.

An open-weight model is a model whose trained weights anyone can download, so we can run it on our own hardware or rent it from any provider. These are some of the strongest for code, as of October 2026:

• Kimi K3 (from Moonshot AI).
• DeepSeek V4 (from DeepSeek).
• Qwen3.8 (from Alibaba).
• gpt-oss (from OpenAI).

A small open model runs on our laptop with a tool such as Ollama. The largest ones need data-centre graphics cards, so most of us would rent them by the token through a service such as OpenRouter.

To code with one, we also need an agent that can talk to it. Several agents are open source themselves: OpenCode, Cline, goose, OpenHands and Aider. OpenAI’s Codex and xAI’s Grok Build publish their code too, though each is built around its maker’s models.

Here is what that route gives us, and what it costs us:

• Our code can stay on our own machine, which matters when a client forbids sending it to a cloud.
• There is no subscription. A machine strong enough for a big model costs real money, though, and a rented model bills us for every token.
• A model small enough for our laptop is far weaker at coding than the large models behind the paid agents.

Claude Code works only with Claude models. We can reach them through Anthropic, Amazon Bedrock, Google Cloud or Microsoft Foundry, and Anthropic does not support pointing Claude Code at any other model.

Why This Course Uses Claude Code

In our opinion, Claude Code is the best agentic coding tool available right now, and many developers agree. The 2026 Stack Overflow Developer Survey, with more than 30,000 answers, put Claude Code first among the coding agents developers use, at 66 percent.

The Pragmatic Engineer’s 2026 survey of more than 900 engineers found it the most used AI coding tool. It was also the most loved, with 46 percent against 19 for Cursor and 9 for Copilot.

Benchmark scores are much closer, and the top agents trade first place from month to month. In our daily work, we notice other things more: how an agent reads a large codebase, how it asks before a risky step, and how well it follows our instructions. That is where Claude Code serves us best.

We can now tell an agent from a chat, say where Claude Code runs, and know what it costs us to start. We can also say what a token is, and place Claude Code beside Claude chat, Cowork, the other coding agents and the open models. The new terms in this lesson are agent, token, agentic coding and open-weight model. In the next lesson, we install Claude Code.