Computer use vs connectors: when Claude should click, and when it should call a tool
Why the best Cowork workflows start with the most structured route, where the browser actually fits, and how to keep computer use from turning into cleanup
If Claude is clicking through Slack on your desktop while the Slack connector is already enabled, you’ve probably chosen the wrong route.
That’s the mistake I think a lot of Cowork users are about to make.
Computer use looks like the advanced option because it’s visible. You can watch Claude move through windows, click buttons, open apps, and work across your machine. That makes it feel like the most capable path.
Anthropic’s own docs describe a different order. In Cowork, Claude is supposed to use the most precise tool first: connectors, then browser, then screen interaction. Anthropic also says connectors are the fastest and most reliable path, while screen-based work takes longer and is more error-prone.
That changes the whole mental model.
Computer use isn’t the default because it can do more things. It’s the fallback when the task actually depends on the desktop.
A screen is a noisy place to work. Windows move. A tab opens in the wrong place. A modal covers the field Claude was about to use. The app state depends on whatever happened five minutes earlier. When you use the screen for work that could’ve gone through a connector, you’re choosing the messiest layer for no real gain.
Connectors are for work with a known shape
Connectors make the most sense when the task already maps to a predictable action.
That includes work like pulling context from Slack, finding files, drafting a message, updating a project, or reviewing a design without forcing Claude to visually navigate the whole interface. Anthropic’s current interactive connector docs make that pretty concrete. The current interactive connector list includes Amplitude, Asana, Box, Canva, Clay, Figma, and Hex. The computer use routing doc also uses Gmail, Google Drive, and Slack as examples of the connector path.
The advantage isn’t that connectors feel more enterprise. The advantage is that Claude can work closer to the data and farther from the interface.
It doesn’t need to visually parse the whole app. It doesn’t need to infer which menu matters. It doesn’t need to click its way through a layout just to reach the thing it already knows how to do. For operator work, that usually means fewer retries and less cleanup.
The browser is its own layer
This is the part people flatten too quickly.
Cowork does not jump from connectors straight to taking over your desktop. Anthropic explicitly puts the browser in the middle. When there isn’t a connector for the tool you need, Claude can navigate the Chrome browser to work on the task using Claude in Chrome. Claude in Chrome itself is available in beta on paid plans.
That matters because a lot of real work isn’t local-app work. It’s browser work.
Internal dashboards. CMS panels. Analytics views. Admin consoles. Vendor portals. Back-office tools your team uses every day that don’t happen to have a connector.
Those jobs need access to the web surface. They don’t need blanket control of your whole machine.
That’s why treating every non-connector task like a computer-use task is too blunt. A lot of the time, the browser is the better fit.
What actually belongs on the desktop
Computer use starts making sense when the interface itself is the constraint.
Anthropic describes it as Claude directly interacting with your screen by clicking, typing, and navigating desktop apps. It can also work in the browser, open files, and run dev tools. Anthropic’s examples and guidance make the intended use pretty clear: direct screen interaction is for the cases where connectors and browser routing don’t get the job done.
That gives you a practical boundary.
I’d use screen interaction for:
desktop-only software
local file workflows
awkward internal tools with no sane export path
cross-app sequences that really do live on the machine
I wouldn’t use it just because it looks more agentic.
That’s the trap. Visible motion gets mistaken for better workflow design.
Where I’d keep it on a short leash
Anthropic’s safety guidance is unusually direct here.
Cowork is a research preview with unique risks. Anthropic says Cowork activity is not captured in audit logs, the Compliance API, or data exports, and explicitly says not to use Cowork for regulated workloads. For computer use specifically, Claude takes screenshots to understand what’s on screen, can see visible information in the apps you’ve allowed, asks permission before accessing each application, and runs outside the virtual machine Cowork normally uses. Anthropic also advises against using computer use for sensitive information, including financial, legal, medical, and other personal data. Some sensitive apps, including investment, trading, and cryptocurrency apps, are blocked by default.
So I wouldn’t start here:
moving money
handling contracts
working inside healthcare or HR systems
deleting or restructuring important files
taking customer-facing actions I’d hate to explain later
That doesn’t make computer use weak. It just means the boundaries matter.
The first workflow I’d trust
The first useful Cowork workflow will usually be mixed.
Say you want a morning brief.
Claude pulls context from connected tools. It grabs files through the connector layer where possible. It opens a browser-only dashboard if the metrics live in a web tool without a connector. Then, only if needed, it touches the desktop for one blocked step involving a local file or app. The output is a memo or packet that a human reviews before anything gets sent or changed.
That pattern makes more sense than handing Claude your machine from step one.
Each layer is doing the kind of work it’s actually good at. Connectors handle structured retrieval and direct actions. The browser handles web tools that sit outside the connector catalog. Screen interaction handles the ugly last mile.
That’s the version I’d trust first.
Anthropic’s own guidance points in that direction too. Their Cowork safety docs tell users to avoid sensitive local files, stay cautious with browser access, use trusted sites and tools, and monitor Claude for suspicious actions or prompt injection.
