xAI rolls out Grok Computer beta

xAI has begun beta testing “Grok Computer,” an AI agent Elon Musk said can carry out office tasks on a computer, with a broader beta planned within days. (aibase.com) The product surfaced after users spotted an `enable_grok_computer` flag in Grok’s web interface code in March, and Musk replied on X that it was “coming out soon.” On April 13, he said the beta had started and that a wider beta would follow in three days. (inshorts.com) (aibase.com) A computer-use agent is software that works a screen the way a person does: it reads screenshots, moves a cursor, clicks buttons, and types into apps. OpenAI describes its Operator system that way, and Anthropic says its computer-use models can navigate a desktop through mouse and keyboard actions. (openai.com) (anthropic.com) That puts xAI into a race that already includes OpenAI’s Operator, Anthropic’s computer-use tools, and Google’s Project Mariner. Google told users in May 2025 that Mariner was a web-browsing agent, while TechCrunch reported these systems were still experimental and prone to mistakes. (openai.com) (anthropic.com) (techcrunch.com) xAI’s pitch is workplace automation inside the Grok ecosystem rather than a standalone chatbot answer box. The company’s own product pages already describe Grok as a tool for “deep work,” coding, document creation, real-time search, and agentic tool calling. (x.ai) (docs.x.ai) The timing also fits Musk’s effort to tie together models, data, and distribution across his companies. In March 2025, xAI and X were merged in a transaction that valued xAI at $80 billion and X at $33 billion, and xAI’s site later said SpaceX acquired xAI on February 2, 2026. (sullcrom.com) (x.ai) xAI has been building toward more autonomous products for more than a year. Its February 2025 Grok 3 launch framed the model as a “reasoning agent,” and current xAI documentation says newer Grok models support function calling and agent tools through the application programming interface. (x.ai) (docs.x.ai) What xAI has not published yet is the safety documentation, task limits, or pricing that rivals have attached to similar products. OpenAI released a system card for Operator, and Anthropic said its computer-use beta was “experimental” and “error-prone” when it launched. (openai.com) (anthropic.com) If the wider beta arrives on schedule this week, the next test will be whether Grok can do the dull parts of office work reliably enough for people to trust it with a live screen. (aibase.com)