Anthropic ships three products, and they all run the same models. What changes is the harness around the model, how many tools it has, and how much control you keep. In this course we call them Chat, Work and Code. Anthropic’s names are Claude (the app), Claude Cowork and Claude Code.

Chat and Cowork are becoming one (September 2026)

Since 16 September 2026 Anthropic is merging Chat and Cowork into one, called simply “Claude”. It comes to Pro and Max plans first. Team and Free plans follow soon, and Enterprise plans get at least 30 days’ notice. The Code tab stays where it is. If your account has not moved yet, the desktop app still shows three tabs: Chat, Cowork and Code. If it has moved, you start a task and connect folders in the same “Claude” conversation where you chat.

We keep the three names in this course because they are three levels of tools. Chat mostly talks, and runs small programs in a sandbox in the cloud. Work does whole tasks in a sandbox, with the folders you connect. Code works on your own computer. That picture stays true even when two of them share one window. See how a sandbox works.

About the names: the product is “Claude Cowork”, or “Cowork” for short. “Work” is only our name for it in this course. Nothing is called “Claude Work”. ChatGPT has a “Work” mode too, but that is a different product from a different company.

ChatWork (Cowork)Code (Claude Code)
AnalogyAn advisor on the phoneA contractor you hand a folder and a taskA colleague at the next desk with your keyboard
WhereWeb, desktop and mobile appDesktop app; web and mobile in betaTerminal, VS Code, JetBrains, desktop app, web
Sees your files?Only what you paste or uploadThe folders you connectThe project folder you open, and anything else you allow
Where the work runsAnthropic’s serversAnthropic’s servers by default, in a sandbox made for each session. It reaches your folders through the desktop appYour own computer
ToolsFew: web search, connectors you enable, small programs in a sandbox in the cloudFiles in connected folders, programs in its own sandbox, connectorsEverything: files, the programs and logins on your computer, web, connectors, plus your own skills and automations
ControlYou edit the answer yourselfYou choose the folders and connectors and an approval mode: approve each step, approve automatically, or skip approvalsYou set a permission mode and approve individual actions, or set rules for what runs on its own
Best atThinking, drafting, quick questionsFinishing a self-contained task in a folder without a terminalAnything that needs many tools, repetition, sharing with colleagues, or automation

When to use which

  • Chat when there are no files involved and you want to think out loud: a question, a draft, a comparison, a translation. Improve the answer yourself, step by step.
  • Work when the task lives entirely in a folder and you want a finished result without opening a terminal: “go through these 40 invoices and make a table”, “tidy these meeting notes into one summary”. You set the boundary (which folders, which connectors) and how often it should ask before acting.
  • Code when you want the full toolbox: several folders, programs, the web, connectors, saved instructions, repeatable skills, and work you share with colleagues through folders and Git. Also when you want to see and approve each step.

What each one can reach

The model is the same in all three. The gear is not. This is the table from lecture 03.

ChatWorkCode
Connectors from claude.ai✓✓✓
Runs programsIn a sandbox in the cloudIn its own sandboxOn your computer, as you
Your installed programs and logins, such as gws✗✗✓ every one
An API with your own keyHard: the key would go into the chatHard: the key would go into the chat✓ from a script on your computer
The screen (computer use)✗✓ Pro and Max plans✓ Pro and Max plans. On Windows only in the desktop app
  • All three get connectors. Switch one on at claude.ai and it shows up in Chat, Work and Code, because they share your account.
  • Only Code uses the programs on your own computer. A command-line program you installed and logged in to, like gws, works in Code. Work can install programs too, but only inside its own sandbox. That sandbox cannot see the programs on your computer, and it reaches only the websites on its allowed list. Chat runs small programs in a sandbox in the cloud.
  • On a company Team plan an owner decides which connectors you can switch on, and none of the three can use the screen.

Do we need Work at all?

Fair question. Work is the same engine as Code, with a simpler window around it. It exists for people who will not open a terminal, and for tasks that fit neatly inside one folder. If you finish this course you will be comfortable in Code, and Work becomes a convenience rather than a necessity: pick it when working inside one folder and the desktop app make a task simpler, pick Code when you need more.

Where the work happens. Code runs on your computer: the tools act on your disk, in the folder you opened. Work runs in Anthropic’s cloud by default. Each session gets its own sandbox on Anthropic’s servers, and when it needs a file from your folder it asks the desktop app on your computer for it. So the files it works on pass through Anthropic’s servers. For most work that is fine, but it is worth knowing when a folder holds something your company is careful about. A local mode also exists, where the work runs in a closed box on your own computer (a virtual machine). Either way, Work can read and change the real files in the folders you connect.

Why the course aims at Code: it shows the machinery. Once you have watched an agent read a file, ask permission and write a result, Chat and Work stop being mysterious. They are the same thing with fewer tools.

Sources