Cowork

Cowork is an assistant for your files. You share a folder with it, describe what needs to happen, and it works there: reading documents, sorting them, writing summaries, filling in a spreadsheet, building a presentation.

Everything happens on your device. The folder never leaves it as a whole, only what the assistant actually reads goes to the model.

What it can do

Area	Examples
Find things	Search the contents of every file in the folder, find files by name pattern, look up metadata
Read	Text, Markdown, CSV, PDF, Word, Excel, PowerPoint, images, and inside ZIP, TAR and TAR.GZ archives
Scanned PDFs	A PDF with no embedded text is put through text recognition automatically, and the result says so
Calculate	Sums, counts, averages, minimum, maximum and median over a spreadsheet, grouped and filtered, computed on your machine instead of guessed from text
Look at things	An image is passed to the model as an actual image. A PDF page or a slide can be rendered so the assistant sees the layout it produced
Write	Create and edit real Word, Excel, PowerPoint and PDF documents, plus plain text, Markdown and CSV
Tidy up	Move, copy, rename, create folders, delete
Remind you	Create, list and cancel reminders

Things people actually use it for

• "Sort the invoices in this folder into subfolders by year and supplier, and list anything you could not classify."
• "Read all twelve interview transcripts and write me a summary per interview plus one across all of them."
• "Fill column F of the price list with the margin, and tell me which rows have no purchase price."
• "How much did we spend per cost centre last quarter according to this spreadsheet?"
• "Turn this Word document into a presentation in our corporate template."
• "Rename all these photos to the date they were taken."

Sharing a folder

The assistant works exclusively in folders you share. Without a folder there is nothing to do, and Cowork says so rather than guessing.

If no conversation is open, Cowork shows a start page with your sessions and due reminders. They are the same sessions as in the sidebar history, most recently used first. A session that waits for you or has failed says so next to its name.

Every file access is checked against those folders. A path outside them is rejected by the app itself, whatever the model asks for. You can widen that under Data storage, and it is off by default for good reason.

Adding a folder later

Before the first message, the folder button below the composer does it. During a running conversation, the same menu sits on the Shared folders button at the top of the Files pane: Share another folder adds one, for example a template folder next to the folder being filed.

Taking a folder out again

There are two ways, and they reach differently far:

• Remove from this chat in the folder menu takes a folder out of this conversation only. It stays shared. A conversation's primary folder cannot be removed this way.
• Stop sharing is in the right-click menu of the folder's root row in the file tree. After a confirmation the folder is taken away from every conversation, its running processes are stopped, and its permissions are revoked.

Below the composer

As in Code, the bar below the input field holds + for attachments, Tools (web search, image generation, which integrations are reachable), the model menu with its Effort submenu, and the mode. Model and effort are remembered per conversation. On request the assistant saves a generated image or a diagram from the conversation as a file (PNG or SVG) in a shared folder. Special app in the header binds a special app to the conversation before the first message.

Permission modes

The mode decides how much the assistant may do without asking. It can be switched at any time, also with Cmd/Ctrl+Shift+M.

Mode	Behaviour
Plan mode	Reads only. Produces a plan, changes nothing
Normal	Asks before every change, every deletion and every command
Accept edits	File changes run without asking. Commands and external tools still ask
Auto	Everything runs without asking

Before Accept edits or Auto takes effect, the app spells out what that means and asks you to confirm. Plan mode and Normal switch immediately.

What Auto really means

Auto is not "the assistant is more capable now". It is the same assistant with the questions turned off:

• Files are written, overwritten, moved and deleted without a separate confirmation.
• The file tools stay inside the shared folders. That boundary is enforced by the app, not by the mode.
• Commands run on your machine here too, and in Auto they run without a confirmation, critical ones included, such as anything under sudo or a downloaded script piped into a shell. In Normal and Accept edits a critical command always produces an approval card. Only the interactive terminal is a Code feature.
• A sub-agent still asks in Auto when it wants to run a critical command or touch a path outside the shared folders, because you did not watch it decide on that call.
• Commands are only available if your account may also use DeutschlandGPT Code. Without that permission the assistant works with the file and document tools only, and nothing on this list about commands applies.
• Actions of connected services keep their own gate unless you also switch on Approve connections automatically.

Use it in folders where an unwanted change would not hurt, for example a working copy or an export folder. For your only copy of something, Accept edits gives you almost the same flow while a deletion still stops and asks.

Approvals

When the assistant wants to change something, an approval card appears showing what would actually happen: a diff for a text file, the affected cells for a spreadsheet, the size of a deletion.

Key	Action
Cmd/Ctrl+Enter	Approve
Cmd/Ctrl+Shift+Enter	Approve everything pending
Esc	Decline
Alt+1 to Alt+9	Select a card when several wait
Alt+Up / Alt+Down	Move the selection

Each card can also grant a narrower standing permission. Always allow binds to one kind of action in one shared folder, which is why the button names the folder. The Permissions menu in the header and Granted permissions in the Activity pane list every permission you granted and take a single one back, or all of them at once. When you stop sharing a folder, its permissions go with it.

Permission to delete is deliberately never remembered beyond the current conversation.

Actions of connected services, so integrations and MCP servers, have their own gate on the server side. They can be approved automatically with Approve connections automatically, independently of the file gate. Integrations covers what those actions are.

Talking to a running assistant

You do not have to wait for a turn to finish.

• Send while it works. Your message goes into the running turn at the next step boundary, so it steers the work in progress instead of arriving after it. The exception is a pending approval or a slash command, which wait for the current step.
• Several messages typed in a row are queued and shown as such, then dispatched in order. Send now on a queued message stops the current turn and sends it right away.
• Stop ends the turn. Everything already done stays done, the assistant does not roll back.
• If a turn hits its step budget with work left, it does not just stop: a card asks whether it should keep going, and Keep working resumes the same turn with every step so far intact.

Sub-agents and background tasks

For larger jobs the assistant can hand work to a sub-agent: Explore only reads and returns an answer, Agent carries out a self-contained sub-task. Each has its own row in the conversation, and every tool it calls passes the same approval gate.

What keeps running next to the conversation, a sub-agent in the background or a command that takes longer than 30 seconds, is listed in the Background tasks pane. There you can Stop it, copy its output, and tidy up with Remove finished.

Slash commands

Typing / in the composer opens the list.

Command	Effect
/help	Lists every available command
/clear	Starts a new conversation
/model	Switches the model. Pass part of a name, or nothing to see the current one
/effort	Sets the reasoning effort: low, medium or high
/plan	Toggles plan mode, optionally with the task right away
/compact	Summarises the conversation so far to free up context
/instructions	Looks at the folder and writes standing instructions for it
/integrations	Opens the connected services
/mcp	Opens the integrations, filtered to MCP servers
/usage	Opens your usage and limits

/anweisungen still works as an alias for /instructions. /effort is only listed when the current model supports a reasoning effort.

The task list

For a job with several steps, the assistant keeps a list of its steps, each with its state. As soon as there is one, a Plan icon appears at the left of the panel icons in the header and opens the list as a pane. When the list changes while the pane is not visible, the icon gets a dot.

Switch it off with Show task list under Chat settings. With it off, the assistant works without a list rather than writing a stand-in into the conversation.

Standing instructions

Some things you would otherwise repeat in every conversation: how invoices are named here, which folder is off limits, that summaries should be in English. There are two places for them.

• In the folder: the file .deutschlandgpt/anweisungen.md in the shared folder. /instructions has the assistant look at the folder and create the file, and afterwards you edit it like any other file. It applies to every conversation in that folder, including colleagues working in the same shared drive folder.
• In the project: a project's ⋯ menu in the history has Instructions. These are stored on your account and apply only to conversations in that project, on top of the folder's instructions.

Where the two disagree, the more specific one wins. Whatever you say explicitly in the conversation always takes precedence.

The assistant reads the folder file first and the project's instructions after it. Where they disagree, the project's instructions win.

History and projects

The Adjust view menu above the history filters by Last activity, groups by date, folder, status or Custom projects, and sorts in Custom order, by name, creation date or last activity. By default the history is grouped by Custom projects in Custom order: a conversation stays where it is when you open it or write in it, a new one starts at the top, and you reorder by dragging a conversation above or below another. Pinned, open and running conversations stay visible under every filter, and Reset restores the default view. When grouped by folder, each folder group has a + for a new conversation in that folder.

Projects are your own way of sorting: New project creates one, and you drag conversations onto it or use Move to project. Your account remembers a project's folders. On another device the app offers to share them and never shares them silently.

A conversation's menu offers the same entries as in Code, such as Pin, Archive, Mark as unread, Output style and Keep computer awake. Overview and history describes them.

Opening a conversation switches the workspace to that conversation's folders. If a folder has been renamed, moved or unshared, the conversation still opens read only and offers to pick a folder again, which rewrites the assignment. The history is never lost with the folder.

Sharing a conversation

Click Share Conversation in the header, then Create link. People in your organization can open the link. It shows the conversation as it is at that moment, and your shared folders do not go along. What the assistant read or changed is visible only as far as it appears in the history.

Two people cannot work in the same conversation. With Continue chat your colleague gets their own copy as a normal chat, without access to your folders. To keep working on the same files, they share the folder on their own device, for example a shared network drive.

The Activity pane

The side panel's Activity pane is where you watch the work and take permissions back. It shows every file access live while the assistant runs, so you see what it reads and writes as it happens rather than afterwards. Under Granted permissions it lists the standing permissions of this workspace, each revocable on its own, plus Revoke all. The clock icon in the pane's header opens the change history.

Undoing a change

The Change history in the Activity pane has two tabs: Changes holds saved earlier versions of every file the assistant changed, Trash holds what it deleted. Bring back a single entry with Restore this version, or roll back everything one request changed with Restore the state from before this request.

That history covers earlier conversations too, since it belongs to the folder rather than to the chat. It is kept for 30 days and up to 1 GB, whichever comes first. Clear history and Empty trash throw it away earlier, each behind a confirmation, and neither can be undone.

More