Claude's Computer Use is coming to Cowork and Code

Claude's Computer Use will soon be available for Cowork and Code as well. This means Anthropic's AI chatbot can also operate a computer in these two modes, just like a human would – opening documents, using the browser, and developer tools.

Computer Use itself is not entirely new. The feature already existed in the regular chatbot. Now it is being expanded. Initially, all users with a Pro or Max subscription will have access. Another restriction: Computer Use in Cowork and Code will only work in the macOS desktop app. Windows is expected to follow soon. Of course, this is an “early research preview” – a disclaimer all AI companies use, as AI applications are prone to errors.

Claude: Connector before Computer Use

Before Claude accesses Computer Use, the chatbot attempts to utilize so-called Connectors. These are direct connections to other services such as Gmail, Google Drive, Slack, and more. Only when no connector is available does Computer Use start. Claude can then operate the browser – according to the report, this is apparently limited to Chrome. Additionally, Claude can use everything that is visible on the screen. To achieve this, the chatbot takes screenshots, which the AI model then processes.

Anthropic also explains that Claude will always use the fastest solution. In the support article, the company gives an example that sending a message on Slack via a connector takes only a few seconds, while Computer Use would take significantly longer. Before Computer Use is activated, Claude always asks for permission.

Similar functions are also available from other providers. OpenAI, for instance, has the Operator and ChatGPT Agent in its arsenal. Both are agentic capabilities that can also use a computer, just like a human would. Microsoft's Copilot can perform so-called Actions – these are tasks that are completed in the background and require, for example, filling in input fields or similar, for which the AI model needs to process a website or understand other applications. In Copilot Studio, there are also tests with the AI function also called “Computer Use”. Google has been testing “computer usage” by Gemini for some time.

Risks of Computer Use

All providers share the commonality of warning about its use. According to Anthropic, Claude is trained to recognize risky tasks. This is said to include prompt injections. These are hidden instructions to AI chatbots that cause unwanted behavior for the actual user. However: Even Claude may not be able to detect all attack attempts under certain circumstances.

Since “Computer Use” takes screenshots, Anthropic also warns that one should pay attention to all visible information. In principle, Claude should not execute stock trades or investment transactions, never enter sensitive data anywhere, and not capture images of faces.

Anthropic also points out that no banking details should be provided. Cowork also has a memory, meaning what has once been communicated to the chatbot remains stored. Cowork does not run in a secure environment with Computer Use.

(emw)