AnthropicのClaudeがコンピューターを操作できるようになる

2 Min ReadSmith Collection/Gado via Getty ImagesThe generative AI vendor has given its Claude large language model the ability to take control of your computer and perform certain tasks on request, as it looks to rival personal agent framework OpenClaw.The update, revealed in a March 23 blog post, enables the Claude Code and Claude Cowork tools to interact with a user’s machine, opening files, working with the browser, running tools and filling in spreadsheets.Once prompted, Claude will look to connected apps such as Slack, Calendar or other integrations first, but if there is no connector for the tool needed, it will work locally on the device, pointing, clicking on and navigating to what’s on screen to do the requested task itself -- although it always asks for permission first, according to the vendor.A video released by the vendor, posted on YouTube, demonstrated the new possibilities Claude offers.Among the examples it showcases is a user sending a request from their phone to their computer to export a pitch deck and then attaching it to the invitation for a scheduled meeting later that day.Related:OpenAI Updates Agents SDK, Aims at Building Secure AgentsClaude springs to life on the laptop, executing the task as requested, then messages the user to inform them what specifically has been done.A second use case depicts a user sending a message asking whether Claude can resize 150 photos on their desktop and add logos to each. Again, Claude is shown doing as asked and then sending confirmation.The agent is also able to handle "repeating" tasks if requested, such as scanning emails every morning, plus Anthropic said it is able to remember context across different sessions.However, Anthropic acknowledged that the tech is in its infancy and might not get things right the first time.  “It won’t always work perfectly: complex tasks sometimes need a second try, and working through your screen is slower than using a direct integration. We’re sharing it early because we want to learn where it works and where it falls short,” the blog said.Anthropic also said it has built in safeguards to minimize risk, protecting, for example, against "prompt injection" -- malicious inputs designed to compromise the AI.The updated functionality is available now for Claude Pro and Claude Max subscribers, but only for macOS users, and must be enabled in the desktop app settings.Claude’s update follows the viral success of OpenClaw earlier this year, which showed the huge potential for AI agents to carry out tasks for users on demand, and is likely to pave the way for a wave of competitors from the major AI companies.Related:As AI Infosec Woes Heighten, IBM Intros Autonomous Security ServiceAbout the AuthorContributing WriterGraham Hope has worked in automotive journalism in the U.K. for 26 years, including spells as editor of leading consumer news website and weekly Auto Express and respected buying guide CarBuyer.