In Claude Chat, the assistant answers questions, and the user acts on the response — copying, formatting, and saving the output elsewhere. In Claude Cowork, that dynamic reverses: Claude plans the task, executes it autonomously in multiple steps, and delivers a finished file.
Use Cowork for tasks with one or more of the following traits:
A connected folder gives Claude persistent access to its contents, so it always works with the latest files rather than relying on a one-time upload. Claude can access only folders the user has explicitly connected; everything else on the computer is off-limits.
In a connected folder, Claude works directly with files — creating, editing, organizing, and performing batch operations across an entire collection. For example, Claude can rename a folder of scanned files by date, convert them to PDFs, and sort them into subfolders, all in one request.
A scheduled task runs on a set cadence — such as hourly, daily, weekly, or on weekdays — or on demand, without the user being present to trigger it. This lets Claude automate recurring work, such as a morning briefing or a weekly report, without manual intervention.
Dispatch lets a task assigned in the Claude mobile app be picked up and carried out by the desktop app, using local files, installed plugins, and connected tools already configured in that Cowork setup.
For Dispatch to complete an assigned task, the Claude desktop app must remain open and the computer must stay awake throughout the work. If the desktop is asleep or the app is closed, Claude can’t work on the assigned task.
Computer use lets Claude open applications and interact with them directly — clicking, typing, and navigating menus much like a person would. Because direct connectors are typically faster and more reliable, Claude uses computer use only when no connector or browser-based option is available.
A cross-application workflow combines multiple capabilities within a single task — such as local files, web information, MCP connectors, and desktop applications — to produce a single deliverable, rather than leaving the user to assemble the pieces.
When accessing a tool, Claude always takes the most direct route:
A plugin is an installable, role-specific toolkit that bundles skills, MCP connectors, slash commands, and sub-agents into a single package. Installing a plugin turns Claude into a specialist for that role, automatically applying the appropriate formats, conventions, and tool connections, rather than requiring each component to be configured individually.
Permission modes determine when Claude pauses for approval.
Manually approve — Claude pauses so you can approve each action; best for sensitive work.Automatically approve — Claude runs on its own, pausing only if something seems unsafe.Skip all approvals — Claude never pauses, even for unsafe actions; use only for fully trusted work.Prompt injection occurs when malicious instructions are embedded in files, emails, or web pages to influence what Claude does when it processes that content. The most effective way to reduce this risk is to limit Claude to content from trusted sources.
For complex tasks, Claude automatically delegates work to specialized sub-agents that run in parallel, each handling its own portion before Claude combines the results. For example, in a research task, different sub-agents can investigate multiple sources simultaneously, producing results faster than if each source were investigated individually.
A hook is an automation that runs in response to a specific event during a task, rather than on a schedule. For example, a compliance plugin might define a hook that runs whenever Claude finishes editing a sensitive file.