What Is Claude Cowork? Chat and Claude Code Compared

Claude Cowork is Anthropic's task-oriented workspace for knowledge work. You give Claude an outcome—such as organizing approved files, analyzing a spreadsheet, preparing a briefing, or creating a formatted document—and it can plan and carry out multiple steps with the tools and permissions available in that session.

Cowork is more agentic than a normal chat, but it is not an unsupervised employee. You still choose the accessible files and apps, review permission requests, steer the work, and verify the result.

How Cowork works

A Cowork session can run in Anthropic's cloud and use connected tools, skills, plugins, code execution, and files saved to your Claude account. Some capabilities still depend on the Claude Desktop app:

Local files: Claude can reach only folders connected through Claude Desktop, with the permissions you granted, while the desktop app is open.
Local connectors and plugins: these also require the desktop bridge.
Computer use: on supported desktop systems, Claude can interact with approved apps by seeing screenshots, clicking, and typing.
Mobile continuity: you can open a Cowork session on another surface to monitor progress, answer a question, or redirect the task.

Tasks can take time and may use several sub-agents. Cowork shows progress and lets you intervene rather than requiring you to wait for a single final reply.

Claude Chat, Cowork, and Claude Code compared

Feature	Claude Chat	Claude Cowork	Claude Code
Primary role	Conversation, writing, analysis, and quick assistance	Multi-step knowledge work across approved files, tools, and apps	Agentic software development and technical workflows
Typical tasks	Explain a topic, brainstorm, revise text, analyze an uploaded file	Prepare reports, process documents, analyze data, coordinate connected tools	Inspect repositories, edit code, run commands and tests, review changes
File access	Uploaded or otherwise attached content; available connected sources depend on the chat	Claude account files, connected sources, and approved local folders through Desktop	Files and tools available in its local or cloud development environment
Action model	Mainly conversational, with optional tools	Plans and executes longer tasks with progress, steering, and permissions	Works through development tasks with repository context, commands, edits, and verification
Best fit	A focused answer or collaborative draft	Documents, spreadsheets, research, organization, and recurring knowledge workflows	Codebases, terminals, IDEs, tests, and engineering automation
Where it runs	Web, desktop, and mobile surfaces	Web, desktop, and mobile, with Desktop required for local resources	Terminal, IDE, desktop, web, and supported remote/mobile continuations
Coding knowledge	Not required	Not required for common tasks	Prompts can be plain language, but software-development judgment is important

The boundary is about the work, not simply the interface. Use Chat when the deliverable is an answer. Use Cowork when the deliverable requires several actions across knowledge-work resources. Use Claude Code when the main environment is a software project.

Tasks that fit Cowork

Read meeting notes from an approved folder and create a project plan with source references.
Analyze a spreadsheet, explain unusual rows, and save a reviewed summary.
Collect information from connected email, calendar, or storage services and draft a briefing.
Reformat a set of non-sensitive documents while preserving the originals.
Create a presentation or report from supplied source material.
Run a recurring summary from connected services and deliver the output to a Cowork session.

Start with a task whose inputs and successful output are easy to define. A folder full of mixed, sensitive, or poorly named files is not a good first experiment.

What Cowork can access

Cowork can use only the resources exposed through the current session and your configured permissions. These may include:

files saved to your Claude account;
local folders connected through Claude Desktop;
connectors such as cloud storage or communication services;
installed plugins and skills;
approved websites through browser tools;
desktop applications through computer use on supported systems.

Access is not harmless simply because it is convenient. A read permission can expose confidential information; a write permission can change data; a browser or app permission can trigger external actions.

Permission modes and deletion protection

Cowork provides modes that control when Claude asks before using connector tools. Connector settings may mark tools as always allowed, approval required, or blocked, and organization policies can impose stricter rules. Select the narrowest practical setting.

Anthropic states that Cowork requires explicit permission before permanently deleting files. Treat that safeguard as one layer, not a replacement for backups, version control, or carefully scoped folders. Other changes—such as overwriting content, sending a message, or editing data in an app—may have their own permission behavior.

Computer-use risks

When computer use is enabled, Claude can take screenshots and interact directly with approved applications. Anthropic warns that there is no sandbox between Claude and those desktop apps. Anything visible in an approved app may be exposed to the session, including private information belonging to other people.

Avoid financial accounts, investments, legal documents, medical records, and sensitive personnel data.
Close unrelated apps and documents before starting.
Use per-app permissions and avoid broad automatic approval for unfamiliar workflows.
Do not let untrusted web pages, emails, or documents redefine the task; they can contain prompt injection.
Require a person to approve sending, publishing, purchasing, deleting, or changing access.

How to start safely

Open Cowork on a supported Claude surface and review the plan or workspace availability shown in your account.
If the task needs local files, update and open Claude Desktop and connect a dedicated test folder containing copies.
Describe one output, exact sources, allowed operations, prohibited actions, and how the result should be verified.
Use a read-only first step and inspect the returned filenames or source list.
Authorize a narrowly scoped write only after the source review is correct.
Open the created file yourself and compare it with the sources.
Remove permissions or connectors that the recurring workflow no longer needs.

A practical first prompt

Read the three Markdown files in [approved folder]. Do not modify them. Create a draft project summary with Completed, Blocked, and Next sections. Cite the source filename after every bullet, place uncertain statements under Questions, and save a new file named project-summary-draft.md in [approved output folder]. If that filename already exists, stop. Then read the saved file back and report its exact path.

This prompt names the sources, separates uncertainty, prevents an overwrite, and includes verification. It is more useful than “organize my project files,” which leaves the scope and success criteria unclear.

When Cowork is not the right choice

Use a normal chat for a quick explanation or short collaborative draft. Use a spreadsheet or database tool directly when formulas, data lineage, or exact calculations require deterministic control. Use Claude Code for repository work that needs commands, tests, diffs, and version-control review. Keep high-stakes or irreversible actions under direct human control.

For current product behavior, read Anthropic's Cowork getting-started guide and Cowork safety guidance. Developers can use this Claude Code installation guide, while general users can compare ChatGPT, Claude, and Gemini before choosing a workflow.