Claude Cowork Setup: How It Differs From Chat and Code
A practical guide to choosing between Claude Chat, Cowork, and Code, setting up a safe Cowork task, and checking the result.
Contents
Quick answer: use Chat for a guided conversation, Cowork for a finished outcome that spans files and tools, and Code for software projects, terminal commands, tests, and Git. Cowork is not an upgraded chat window or a simpler name for Claude Code. It is a separate delegation surface with different access boundaries.
The feature and interface details below were checked against Anthropic documentation on August 11, 2026. Availability and labels can change by plan and platform.
Chat vs Cowork vs Code
| Question | Claude Chat | Claude Cowork | Claude Code |
|---|---|---|---|
| What should be produced? | An answer, idea, explanation, or draft | A completed document, spreadsheet, research package, or operation | Code changes, commands, tests, a build, or pull request |
| Who drives the work? | You steer each conversational turn | You define the outcome; Claude plans and executes multiple steps | The agent works inside a development project |
| Main context | Messages, uploads, Projects, connectors | A working folder or Project plus approved connectors and tools | Repository, project files, terminal, Git, and development environments |
| Review focus | Accuracy and direction | Scope, actions, files, and final deliverable | Diff, commands, tests, and repository state |
| Where it runs | Web, desktop, and mobile | Desktop plus web/mobile surfaces with different local-resource access | Terminal, desktop, and web; local and cloud sessions |
Anthropic’s own overview frames Chat as guided conversation, Cowork as delegated complex work, and Code as software creation. A planning question belongs in Chat. Turning several source documents into a finished report belongs in Cowork. Adding report generation to an application belongs in Code.
What Cowork does
Cowork works toward a whole deliverable rather than the next reply. You provide a goal and acceptance criteria; it can gather context from an approved folder, Project, or connector, break the work into steps, execute them, and save the result.
It is useful when three conditions are true: the output is concrete and testable, the inputs live in several files or tools, and the job takes multiple steps. For a single explanation, Chat is simpler. When the work requires a terminal, repository changes, or tests, use Code.
Set up a safe first Cowork task
- Open Cowork and start setup. Select Cowork and enter /setup-cowork in the first session. If the command is not visible, Set up Cowork starts the same guided flow; use Customize afterward to add or manage individual plugins and connectors.
- Create a narrow workspace. Make a folder such as cowork-first-task, add copies of two or three non-sensitive files, and select Work in a folder. Do not grant your whole home directory.
- Specify the result and the check. Name the goal, inputs, output file, restrictions, and acceptance criteria. A good first task is: “Create status-report.md from these three notes, do not modify the sources, and list contradictions.”
- Verify before expanding access. Check which files were used, whether each fact maps to a source, whether the source files stayed unchanged, and whether the action log contains any unexpected sending, publishing, deleting, or external access.
Anthropic’s getting-started guide documents folders and Projects as Cowork workspaces. Its safety guide explains that a task runs in a temporary cloud environment; access to local files is mediated by Claude Desktop and the permissions you granted. If Desktop is offline, a cloud Cowork session cannot reach those connected local resources.
A reusable task template
Goal: [the testable deliverable].
Inputs: use only files in the current working folder.
Output: create a new [filename and format]; do not modify source files.
Restrictions: do not delete files, open external sites, publish, or send messages.
Validation: list the files used, missing data, and the criteria you checked.
Show a short plan first. If required data is missing, stop and ask.
Common problems and checks
Cowork cannot see a file
Check that the file is inside the selected folder, the correct folder or Project is active, and Claude Desktop is running when a remote session needs a local resource. Ask Cowork to list visible filenames without changing anything. If the name is absent, reselect the minimal folder instead of widening access to the entire device.
A connector reads but cannot write
Connectors inherit your account permissions and organization policy. First verify that you can edit the target yourself, then check whether write access is enabled for that connector. Reauthorizing cannot create a permission that the user or administrator does not have.
/setup-cowork does not run
Confirm that Cowork—not Chat or Code—is selected and update the app. Set up Cowork starts the same guided flow; Customize is for adding or managing individual plugins and connectors. Commands from Claude Code do not automatically apply to Cowork.
The task is going in the wrong direction
Stop it, narrow the folder, and rewrite the request around a named output and measurable criteria. “Improve these files” is not testable; “create a new file from these three inputs, preserve the sources, and list gaps” is.
Permission boundaries for the first run
Start with a recoverable folder and backups. Add one connector only when the task needs it, and separate read from write access. Permanent deletion requires explicit permission. Computer use is more sensitive because screen actions do not have the same isolation boundary as file or code operations. Schedule a task only after several successful manual runs; purchases, outgoing messages, and deletion should keep human confirmation.
Where BetterToken fits—and where it does not
Cowork and Claude Chat are Claude user experiences. BetterToken is not a Claude subscription, shared account, or a way to turn Cowork into an API client.
BetterToken is a separate pay-as-you-go API service billed by actual model usage. It becomes relevant after you choose Claude Code and the development workflow needs an Anthropic-compatible API, your own API key, and a Base URL. In that case, follow the BetterToken Claude Code setup guide. Those API settings belong to the Code workflow, not the Cowork interface.
Use one final rule: if you want to steer a conversation, choose Chat. If you want to delegate a multi-step document or service task and review a finished result, choose Cowork. If the deliverable is a tested software change, choose Code. For the first Cowork run, one reversible folder and one new output file are enough.