Definition
What is an AI cowork app?
A chatbot answers you. A cowork app does the work — opening your files, editing them, running the steps, and handing back a finished thing. The distinction is not marketing; it comes down to four specific capabilities that a chat interface does not have.
Reviewed 17 August 2026. Prices are US list, ex-tax, and change often — verify on official pages.
The line
Four capabilities that define the category
A tool needs most of these to be a cowork app rather than a chat window. Use them as a checklist when a vendor makes claims.
It reads and writes real files
Not "paste your text here" — actual access to a folder, a repository or a document library, with the ability to change what is there. Claude Cowork asks you to grant specific folders. Codex works against a repository. Microsoft Copilot reaches your Microsoft 365 content through the Graph.
It takes multiple steps without you
Given a goal, it plans, executes, checks its own output and corrects. This is what "agentic" is supposed to mean, and it is the capability most often claimed and least often delivered.
It can run things
Commands, code, tests, application actions. A tool that can only produce text cannot verify its own work. One that can run tests knows whether its change worked.
It survives your attention wandering
Background or scheduled execution — the task continues when you switch to something else, or when you close the laptop. Codex cloud tasks, Claude Cowork cloud sessions and Perplexity Computer all do this. Microsoft Copilot does not.
In practice
What this looks like on an actual Tuesday
| Instead of this | A cowork app does this |
|---|---|
| Pasting a spreadsheet into a chat window, getting analysis back as text, and retyping it into Excel | Opening the actual .xlsx, adding the columns and formulas, saving it, and telling you what changed |
| Asking for a script, copying it into a terminal, hitting an error, pasting the error back | Writing the script, running it, reading the error itself, fixing it, and running it again until it passes |
| Asking for a summary of a document you paste in, one document at a time | Reading all forty files in the folder, cross-referencing them, and writing a summary document into the same folder |
| Remembering every Monday to pull last week's numbers together | Running that job at 06:00 every Monday whether or not you remember |
Note what all four have in common: the saving in each case is not that the AI is cleverer, it is that the human is removed from the copying and pasting.
Naming
Cowork app, desktop agent, AI agent — same thing?
Roughly, yes, with shades of meaning.
AI cowork app
The friendliest label, and the one Anthropic effectively adopted with Claude Cowork. Implies a non-technical user and work that is not code.
Desktop AI agent
Emphasises that it runs on your machine with access to what is on it, rather than in a browser tab.
Agentic AI
The industry term. Emphasises multi-step autonomy, and gets applied so loosely that it now tells you almost nothing on its own.
AI coworker
Marketing language. Treat any claim built on it with the checklist above rather than the vibe.
FAQ
Cowork app FAQ
Is ChatGPT an AI cowork app?
The ChatGPT desktop app now is, in part — since July 2026 it contains Codex and Work modes that operate on files and repositories. Plain ChatGPT chat is not; it produces text about your work rather than doing it.
Do I need to be technical?
Not for all of them. Claude Cowork, Microsoft Copilot and Perplexity's Personal Computer are built for plain-language instructions. Codex genuinely is a developer tool and assumes you understand repositories and diffs.
Do they work offline?
Mostly no — the models run in the cloud even when the app is on your desktop. The exception is a local-model tool like Msty, which runs models on your own hardware via Ollama, MLX or llama.cpp and can work fully offline. The trade-off is that quality and speed are bounded by your machine.
Are they safe to point at my real files?
With care. Every one of these sandboxes to some degree — Claude Cowork to folders you grant, Codex to a workspace with network off by default, Microsoft's Copilot Actions to four named folders. But all of them can modify and delete inside their permitted area, and both Anthropic and Microsoft state in their own documentation that you carry responsibility for what the agent does. Start on copies.
What is prompt injection and should I care?
It is when content the agent reads — a webpage, a document, an email — contains instructions that hijack it. Anthropic says the risk is "non-zero" even with its mitigations, and Microsoft documents it in its own Windows security guidance. Yes, you should care, particularly if the agent can send email or touch money.