Cowork: turn any folder into an AI project workspace

What Cowork is for #

Cowork is the default for questions and work with files: a thesis, a research project, a report or a folder of notes. Start without a project for a quick question, or choose a folder and give the project’s file tools that workspace. Library, browser and connector tools use their own access settings and may provide additional sources.

Every conversation is saved and grouped by project, so a long-running piece of work keeps its full history. Signed-out Cowork work is not synchronized. After sign-in, Cowork sync is enabled by default unless you turn that category off; file E2E is a separate optional setting and starts off.

Working with your files #

Ask the agent to read notes, restructure a draft or turn findings into an outline. Follow the file reads, tool activity and visible diffs. File and command actions depend on the available tools, access mode and approvals; deletion permission can be remembered for the task or session, so a new confirmation is not promised for every file.

Attachments become real files: drop a document into the conversation and it is written into the project folder, so the agent works on the actual bytes rather than a copy that drifts.

Real documents out: Word, Excel, PowerPoint #

With “Run code” enabled for a project, the agent can generate real .docx, .xlsx and .pptx files — an academic report with numbered headings and a reference list, a budget spreadsheet, a slide deck. Generator scripts run in an isolated process, scratch files are cleaned up automatically, and the finished file lands in your folder with an “Open in Word/Excel/PowerPoint” card at the end of the conversation.

Skills: teach the agent your formats #

Agent Skills are reusable instruction folders the agent loads on demand — the bundled ones teach it how to produce proper documents, spreadsheets and presentations, and you can author your own in Settings → Agent Skills (a skill is just a SKILL.md file with instructions).

A skill costs almost nothing until it’s used: the agent sees a one-line description and loads the full instructions only when the task calls for them.

Instructions, memory and the right rail #

Each project carries its own instructions (tone, formatting, rules) and its own memory file the agent maintains across conversations. The right rail keeps both one click away, alongside the project’s files and scheduled tasks.

Frequently asked questions #

Can Cowork edit my files without asking?

Edits may proceed under the permissions you selected and appear as diffs. Project file tools operate in the chosen workspace, while Library, browser and connectors have separate access. Deletion approvals can cover the task or session; review those permissions before starting.

What document formats can Cowork generate?

With “Run code” enabled it produces real .docx, .xlsx and .pptx files — reports, spreadsheets and slide decks — saved straight into your project folder, ready to open in Word, Excel or PowerPoint.

What are Agent Skills?

Reusable instruction folders (a SKILL.md file) the agent loads on demand. Bundled skills teach it to make proper documents; you can author your own in Settings → Agent Skills.

Can the agent touch files outside my project folder?

Project file tools are scoped to the chosen workspace. That does not limit every tool to that folder: enabled Library access, browser sessions and connectors can reach their own permitted sources. Review each tool’s access and the task’s file and command permissions.

Where do generated documents end up?

Right in your project folder. When the agent finishes a .docx, .xlsx or .pptx it adds an “Open in Word/Excel/PowerPoint” card at the end of the conversation, and the real file is already saved alongside your other project files.