Claude’s Computer Use feature is now live: it can look at the screen, click the mouse, and even type

Claude has recently taken a big step beyond “just chatting” by adding a public beta “computer use” capability. Put simply, you can have Claude view the screen, move the cursor, click buttons, and enter text—completing tasks the way a person operates a computer. This article explains Claude’s new feature from an editor’s perspective: what it can do, how to use it, and who it’s for.

What exactly has been updated in Claude “computer use”

The highlight of this update is that Claude now offers a “computer use” capability on the API side, allowing developers to direct Claude to navigate UIs. Claude makes judgments based on what’s on the screen and then performs actions such as clicking, typing, and jumping between pages, chaining together steps that previously required manual work. It’s worth noting that Claude officially states this is still in an experimental stage, and it may occasionally lag, click the wrong thing, or behave inconsistently across steps.

Where can you access this capability

At the moment, “computer use” is available in beta via the Anthropic API, making it easier to integrate Claude into automation workflows or internal tools. Claude also supports building similar capabilities on Amazon Bedrock and Google Cloud Vertex AI, which should make enterprise deployment smoother. Meanwhile, the upgraded Claude 3.5 Sonnet is already available to all users, with a particular emphasis on improved coding performance.

Which real-world scenarios is Claude suited for: turning “dozens of steps” into “one sentence”

When a task involves lots of repetitive operations, Claude’s value is most obvious—for example, entering items one by one in a web admin dashboard, bulk-filling content in forms, or collecting information across multiple pages and then pasting it back in. You can also have Claude follow a step-by-step “checklist,” turning easy-to-miss clicks and inputs into a fixed process. Some teams are already exploring having Claude execute chained tasks spanning dozens to hundreds of steps for product evaluations, automated verification, or internal operations tools.

Pitfalls and boundaries to note before using Claude computer use

Because Claude “looks at the screen and then acts,” changes in screen state, pop-up overlays, and differences in button styling can all cause misjudgments, so it’s best to provide clear steps and error-tolerant instructions. For account logins, payments, or sensitive data, it’s recommended to require human confirmation for key steps and to enforce least-privilege access controls. On the security front, related updates have undergone pre-deployment testing in collaboration with U.S. and U.K. AI safety institutes; Anthropic also believes its ASL-2 standard remains applicable, but product teams still need to implement strict risk controls on their side.