Home / Choosing the right way in
Chat, Cowork, Claude Code, three tools, three uses
Each Claude environment answers a different need, none replaces the other two.
Anthropic distributes three distinct interfaces for the same family of Claude models: the chat at claude.ai, the desktop agent Cowork, and the command line agent Claude Code. One word comes up throughout this course and deserves to be defined right away: an agent is a programme that chains several actions together on its own to reach a result you have described, instead of handing you a single answer and stopping there. None of the three is a reduced version of the other two, each answers a different kind of task, and confusing the three wastes time from the very first session.
The same model, different tools
Under the three surfaces runs the same family of models, Opus, Sonnet and Haiku. What sets one surface apart from another is therefore not a different intelligence, but the tools the model can call and the context you supply to it. A chat that answers a question and an agent that reorganises an entire folder use the same brain, with a different set of tools switched on around it. Keep the mental rule that applies in all three cases: a Claude model takes text as input and returns text as output, and the whole art lies in controlling what goes into that input text.
Chat answers, Cowork acts
The chat at claude.ai remains the simplest point of entry: you paste or describe a piece of information, Claude answers within the conversation, nothing else moves on your computer. Cowork changes this logic. It is Anthropic's desktop agent. As recorded on 2 September 2026, it is available on the desktop application for all paid plans, and on the web as well as on mobile for the Pro, Max and Team plans, with an Enterprise account gaining access wherever an administrator has switched it on. These commercial tiers move: the address at the end of the lesson gives the state as of the day you read this. You give it access to specific folders, you describe a result, such as a report written up from scattered notes or an invoice folder sorted by month, and Claude actually reads, edits, creates and organises files on your machine, across several steps, without you coordinating each one.
Claude Code, reserved for repository work
Claude Code runs the same agent engine as Cowork, but Anthropic documents it for a different kind of work: that of teams who write programmes. It reads the whole set of files in a software project, edits them, runs commands, and handles the moves specific to development work. These moves carry names that will not serve you here, and this course does not teach them: if nobody in your organisation writes programmes, Claude Code does not concern you. The official distinction fits in one simple sentence: code work, fixing a bug, opening a merge request, running tests, goes to Claude Code, knowledge work, drafting a document, organising files, doing research, goes to Cowork. One nuance is worth knowing before you choose. Cowork's work runs in an isolated environment, often called a sandbox: this protects your computer from a programme misbehaving, but it in no way reduces what Claude can read or do with the access you have given it. Isolation therefore never replaces the choice of the right surface for the task.
Three ways of opening the same request make this clear:
Chat : "Resume cette note de reunion en cinq points."
Cowork : "Range le dossier Factures par mois, sans rien supprimer."
Claude Code : corriger le bug d'affichage du total sur la facture, dans le depot du projet.The first request remains a question put to a party that answers. The second and third ask an agent to act, one on desktop files, the other on a code repository. The next lesson details how to open a first Cowork session, the recommended entry point for anyone who has never used an agent.
Chat, Cowork and Claude Code, four criteria
| Surface | File access | Technical level required | Typical use case | Approval mode available |
|---|---|---|---|---|
| Chat (claude.ai) | No direct access, a file must be copied or attached into the conversation | None, opening a conversation is enough | Draft, analyse, answer a one off question | None, each answer is displayed without acting on anything |
| Cowork | Connected local folders, once access is approved | None, the desktop application, the web or mobile are enough | Organise a folder, fill in a spreadsheet, draft a report from scattered notes | Three approval levels, chosen before execution, from a pause before each action to no oversight at all |
| Claude Code | An entire code repository, read and edited directly | Knowing how to write programmes, and working on the command line | Fix a bug, open a merge request, run tests | Also exists, with its own command line settings |
An employee opens a Claude conversation, pastes her meeting notes into it and gets back a written summary. The next day, she wants the same work done on the year's forty meetings, whose notes are spread across as many files in a folder on her computer. She installs Cowork, gives it access to that folder, and asks it for a summary per file.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: It establishes that the first piece of work required her to bring the text into the conversation herself, and that the second assumes a tool capable of going to read the files where they are already stored.
What this does not establish: It does not establish that the second piece of work succeeded, nor what Cowork can reach outside the folder to which she gave it access.
The three most common miscalibrations
- Too broad Cowork can now act on any file on the computer, connected folder or not, as soon as it is asked to.
- Too narrow This situation shows nothing more than a personal preference for Cowork over the chat.
- Beside the point It establishes that the summary obtained on the first day was of better quality than those requested the next day.
- The three surfaces rest on the same family of Claude models, what changes from one surface to another is the tools available and the context supplied.
- The chat processes the text you bring into the conversation, it does not access any file on your computer by itself.
- Cowork acts directly on connected local folders, without ever opening a terminal, which makes it the entry point suited to a non technical user.
- Claude Code reads and edits an entire code repository and handles commits and merge requests, ground reserved for technical tasks.
- The choice of surface follows the nature of the expected deliverable, not habit nor the surface already open on screen.
Open the three interfaces on your machine, the chat, Cowork if your plan gives you access, and Claude Code if you have it, then match each one to a concrete task drawn from your own work.
These points depend on an interface or a rule that may have changed since this was written. Check them on your own screen before relying on them.
- The existence of approval settings specific to Claude Code, distinct from those of Cowork, is not confirmed in detail by the sources read for this lesson, the corresponding cell remains deliberately general.
- The labels of the three approval levels are served in English in the interface as recorded on 2 September 2026, Manually approve, Automatically approve and Skip all approvals. A label translated or renamed since then is read on screen, not in this course.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Anthropic, getting started with Claude Cowork, surfaces, plans and approval levels consultée le 2026-09-02