Claude Cowork computer use is a research preview, live on Pro and Max plans as of July 2026, that lets Claude open files, run dev tools, point, click, and navigate what is on your screen to finish a task itself, with no setup required (release notes). It is the fix for Cowork's shakiest area: driving apps that had no clean connector.
If you have used Cowork for a while, you already know the pattern. It is great at files and spreadsheets and reports, and then you ask it to do something in a web app it cannot reach, and it stalls. Computer use is Anthropic's answer to that gap. Here is what it actually does, and where the preview edges still show.
What is computer use in Claude Cowork?
Computer use is a mode where Claude operates your machine the way you would. It takes a screenshot of your screen, decides where to click, then clicks, types, and scrolls to move a task forward (support docs). Anthropic's own line is that Claude "can open files, run dev tools, point, click, and navigate to what's on your screen to perform tasks itself, with no setup required" (release notes).
The important nuance is that it does not reach for the screen first. Claude tries the most precise tool it has, in order: a direct connector like Gmail, Google Drive, or Slack if one exists, then browser control through Chrome, and only then falls back to clicking around the screen directly (support docs). Screen clicking is the last resort, because it is the slowest and least reliable of the three. That ordering is the whole design.
This slots into the broader picture in what Claude Cowork is: same desktop agent, now with hands.
Which plans get it, and how do you turn it on?
As of July 2026, computer use is a research preview on the Pro and Max plans only. The support docs are blunt that "Team and Enterprise plans don't have access to computer use at this time" (support docs). That is the reverse of most Cowork features, which usually land on the higher business tiers first, so double-check your own account rather than assuming.
Turning it on is a toggle, not an install:
- Update Claude Desktop to the latest version from claude.com/download.
- Open the desktop app and go to Settings, then General, and find the Desktop app section.
- Switch "Computer use" on.
- Open Cowork (or Claude Code) and start a session.
It runs on both macOS and Windows in the desktop app (support docs). One requirement carries over from the rest of Cowork: your computer has to be awake and the desktop app has to be open for any of this to work. Nothing happens while your laptop sleeps in a bag.
What can it do that Cowork could not do before?
This is the before-and-after, and it is worth being precise about it.
Before this preview, when Cowork could not talk to an app directly, it fell back to driving a browser by clicking around, and that path broke. Through spring 2026, Windows users hit cases where the browser tools stopped registering after a desktop update (GitHub issue), and the Claude in Chrome pairing kept failing inside Cowork sessions (GitHub issue). We wrote up browser automation as Cowork's genuine weak spot in our Cowork use cases post. If a task depended on clicking through a website with no connector, it was the least dependable thing you could ask for.
Computer use is the answer to that specific weakness. Instead of relying only on the Chrome pairing, Claude can now watch the whole screen and click anywhere on it, so an app with no connector and no working browser hook is still reachable. Anthropic's examples include compiling a competitive analysis from local files and connected tools, poking at a phone simulator to spot UX issues, and populating a spreadsheet from several sources that do not share a connector (support docs).
The honest caveat, date-stamped July 2026: this is a research preview, so it is slower than a connector and complex multi-step workflows sometimes need a second attempt. Anthropic says plainly that Claude "can make mistakes" and that working through your screen is slower than a direct integration (Anthropic). It widens what is possible. It does not make screen clicking as fast or as sturdy as a real API. The advice from our use-cases post still holds: reach for a real integration whenever one exists.
Is it safe to let Claude click around your screen?
This is the question everyone asks, and it deserves a straight answer, because computer use is a bigger grant than the rest of Cowork. Regular Cowork tasks run in a sandbox with only the folders you hand over. Computer use is different: there is no sandbox between Claude and your applications when it is driving the screen (support docs). It is looking at, and touching, whatever is on your monitor.
Anthropic built several guardrails around that:
- Per-app permission prompts, so Claude asks before it touches a given app.
- A default blocklist for investment, trading, and cryptocurrency platforms, which you can extend with your own apps.
- Automatic scanning of proposed actions for prompt-injection attempts.
- A stop control you can hit at any moment.
Claude is also trained to avoid stock trading, entering sensitive data, and gathering facial images. Anthropic is candid that "these guardrails aren't absolute" (support docs). So treat it like a capable but new assistant. Start it on apps you trust, keep sensitive data off the screen, and for the first several runs, watch what it does before you walk away. This is the same start-on-a-copy caution we gave for Cowork's file access, applied to your screen instead of a folder.
How do skills combine with computer use?
Here is the part that compounds. Computer use is the hands. A skill is the memory of how the job is done. On its own, a standalone Cowork session forgets everything between runs, so every week you re-type the same long instruction. A skill is a small SKILL.md folder that captures the steps, the format, and the rules once, then loads on demand.
Put the two together and you get something neither has alone. The skill encodes the procedure, "open this tool, pull these three fields, drop them into that template," and computer use executes those steps against the real screen when no connector exists. You describe the workflow one time, and Claude carries it out by clicking, without you narrating it again. A skill like meeting-minutes already structures the output; pairing it with computer use lets Claude also fetch the source from an app it could not otherwise reach.
Keeping it honest: this only pays off where the underlying task is stable and the screen layout does not shift much between runs, since computer use is still preview-grade and can misclick. But when it works, you have built the instruction and the muscle in one place. You write the skill without code using Knack, or start from something close in the marketplace, and Cowork runs it. That is the compounding pitch, caveats included.
FAQ
Is Cowork computer use available on Windows?
Yes. As of July 2026 the preview runs on both macOS and Windows in the Claude Desktop app, once you flip the Computer use toggle in Settings > General (support docs).
Does it cost extra?
No separate charge. It is included in the Pro and Max plans you already pay for. Like the rest of Cowork, though, it is compute-heavy and burns your usage allotment faster than chat, so a long clicking session eats into your limits.
Can it use my browser?
Yes, and it prefers to. Browser control through Chrome sits above raw screen clicking in Claude's tool order, so it drives the browser when it can and only clicks around the screen directly when the browser path does not fit (support docs).
How is this different from Claude in Chrome?
Claude in Chrome is scoped to a browser tab through an extension. Computer use is broader: it screenshots and operates your entire desktop, so it can open native apps, run dev tools, and click anything on screen, not just web pages. In fact browser control is one layer inside computer use, sitting between connectors and last-resort screen clicking.
Can I restrict which apps it touches?
Yes. Investment, trading, and cryptocurrency apps are blocked by default, you can add your own apps to that blocklist, and Claude asks for permission per app before acting (support docs).
Does it work with scheduled tasks?
Yes, through Dispatch. Computer use lets Claude run tasks while you are away, including recurring ones you assign from your phone or desktop (Anthropic). The same rule as all Cowork scheduling applies: your computer has to be awake and the app open when the task fires, or it does not run. If you also work in the terminal, the same capability shows up in computer use in Claude Code, and the driving basics are in how to use Cowork.