# Claude Cowork, Copilot Cowork, and ChatGPT Work explained > What Claude Cowork, Copilot Cowork, ChatGPT Work, and Gemini's agents can do as of November 2026: files they touch, approvals, plans, prices, and countries. **Desktop AI agents** such as Claude Cowork, Copilot Cowork, and ChatGPT Work carry out multi-step tasks in your files, apps, and browser. As of November 11, 2026, Claude Cowork reads and writes in folders you connect on your computer, Copilot Cowork works only in OneDrive and SharePoint, and OpenAI has replaced its former ChatGPT agent with ChatGPT Work and a cloud browser. **Key takeaways** 1. Microsoft says Copilot Cowork cannot access files stored on your device and cannot delete files or folders in OneDrive or SharePoint. 2. Microsoft bills Copilot Cowork for businesses in Copilot Credits, at $0.01 per credit on pay-as-you-go, and offers a consumer preview only in the US and Canada. 3. Claude Cowork asks for explicit permission before permanently deleting any file, whatever approval mode you choose. ## What the desktop agents do A desktop agent is the version of an AI assistant that stops suggesting and starts doing: it sorts the files, drafts the documents, and fills in the web forms. This page is part of [letting AI handle routine tasks](/guides/ai-at-work/automation-and-agents/) and compares the main ones from their makers' help pages as of November 11, 2026, starting with Anthropic's guide to [getting started with Claude Cowork](https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork). We have not tested them. The biggest practical difference is what each one can reach. Claude Cowork works in folders on your computer; Copilot Cowork works in your Microsoft 365 cloud; ChatGPT Work uses a browser on OpenAI's servers. Pick by where your work lives, then check what each asks before acting. ### The four side by side | | Claude Cowork | Copilot Cowork | ChatGPT Work | Gemini | |---|---|---|---|---| | What it works on | Folders you connect, connected apps, a browser; screen control in beta | Outlook, Teams, OneDrive, SharePoint; not files on your device | Connected apps and files, a cloud browser; local files from the desktop app | Connected Google apps; Chrome for auto browse | | Plans | Paid plans: Pro, Max, Ultra, Team, Enterprise | Microsoft 365 Copilot license plus pay-as-you-go; consumer preview on Premium and Pro | Desktop app on every plan; cloud browser on paid plans except Go | Tasks on Google AI Pro and Ultra; auto browse on personal paid plans | | Countries, of our four | All four | Consumer preview: US and Canada | “Supported regions,” not listed | Tasks: not the UK; auto browse: US only | | Before acting | Manual, automatic (the default on most plans), or no approvals; deletions always ask | Approval before sending, posting, or creating, with a risk level | Asks before each new site; confirms before booking or paying | Shows a plan; Tasks always ask before sending, buying, or deleting | | Scheduled tasks | Yes, in the cloud | Yes, and on events | Yes | Yes | ### Claude Cowork Cowork handles long tasks with several steps: organizing files, building spreadsheets with formulas, drafting documents and presentations, and research. Since October 6, 2026, new tasks on Pro, Max, and Ultra run in an isolated session in Anthropic's cloud and keep going with your computer closed; Team and Enterprise can also run them locally. It reads and writes only in folders you connect, and Anthropic suggests a dedicated work folder and backups. You choose between manually approving each action, automatic approval with a safety check on each action, which is the default on Pro, Max, Ultra, and Team, and skipping approvals entirely. Deletions always need your “Allow.” Screen control, in beta on Mac and Windows, has no sandbox between Claude and your applications, in Anthropic's own words. The plans start at $20 a month in the US, and agent work uses up your allowance faster than chat. ### Copilot Cowork Microsoft's agent sends emails, schedules meetings, creates documents, and posts in Teams, with your permissions. Microsoft's [Cowork FAQ](https://learn.microsoft.com/en-us/copilot/cowork/cowork-faq) says it works with files in OneDrive and SharePoint, cannot touch files on your device, and cannot delete files or folders. Before each action, it shows a preview and a button named for the action, such as Send or Post, with a risk level for medium- and high-risk steps. For businesses, it needs a Microsoft 365 Copilot license and pay-as-you-go billing in Copilot Credits, $0.01 each, with consumption that varies by task. In the UK, Claude models are off by default in Copilot, and Microsoft's pages disagree on whether that blocks Cowork. For individuals, Cowork is a preview feature of Microsoft 365 Premium, $19.99 a month in the US, and Pro, in the US and Canada only. ### ChatGPT Work, which replaced agent mode If you are looking for ChatGPT's agent mode, OpenAI's help center says it is no longer available: ChatGPT Work and a cloud browser took its place, and the ChatGPT Atlas browser has been retired. In Plan mode, Work gathers context, asks questions, and proposes a plan you can edit. Its [cloud browser](https://help.openai.com/en/articles/20001280-using-cloud-browser-in-chatgpt) can read, click, and fill in forms, asks permission before visiting each new site by default, and confirms in the chat before a booking or payment; OpenAI notes that allowing a site is not approving an action. ### Gemini and Perplexity, briefly Google's auto browse in Chrome compares products, fills carts, and books appointments on the web, for adults in the US on a personal Google AI plan, and shows its plan before starting. Tasks in Gemini, on Google AI Pro and Ultra, always ask before sending messages, buying, or deleting data. They require Gemini Apps activity to be on, which means your conversations can be used for training, and are not offered in the UK or the European Economic Area. Perplexity Computer runs tasks in a cloud sandbox and charges them in credits: 10,000 a month on Max, none on Pro beyond a one-time bonus. ### Rules the makers give Anthropic advises against scheduling tasks that send messages, make purchases, or touch sensitive files, and Google against scheduling sensitive tasks, because nobody is watching when they run. Anthropic advises against using Claude in Chrome for financial accounts, legal documents, or health information, and recommends manual approval for money, messages, and important files. Anthropic, Google, and Microsoft all say you remain responsible for what the agent does on your behalf. ### What makes it an agent If the difference between these tools and a chatbot is still fuzzy, start with [AI agents in plain terms](/guides/ai-at-work/automation-and-agents/what-is-an-ai-agent/), including what security agencies advise. ### A simpler first step Many routine tasks need a fixed automation rather than an agent. See [the first automations worth setting up](/guides/ai-at-work/automation-and-agents/first-automations/). --- **Prices:** checked on November 10, 2026 on the providers' official websites, in the currency and country shown. A lower monthly price often requires paying for a full year, and promotions change without notice. Amounts may have changed since then.