Detailed Analysis
A user on a MacBook Air M3 running the Claude desktop app under a Pro subscription has encountered a persistent bug in Anthropic's "computer use" feature, where the app is only able to detect Safari as an installed application, ignoring all other software present on the system. Despite granting the necessary macOS Accessibility and Screen Recording permissions, performing a clean reinstall, and repeatedly toggling both the computer use feature and system permissions, the issue persists. The user's troubleshooting steps followed standard remediation practices for macOS permission-based issues, suggesting the problem lies deeper than typical user-side misconfiguration and may point to a bug in how the Claude desktop app enumerates or queries the operating system for installed applications.
This issue is significant because computer use is one of Anthropic's more ambitious and technically complex features, allowing Claude to perceive and interact with a user's desktop environment—taking screenshots, moving the cursor, clicking, and typing—to complete multi-step tasks across applications. For this capability to be useful, Claude must reliably perceive the full landscape of software available on a machine; a failure mode where only one default, first-party application (Safari) is visible severely undermines the feature's core value proposition. Users adopting computer use are typically doing so specifically to automate cross-application workflows, so a bug that reduces functionality to a single browser effectively neutralizes the feature for many real-world use cases.
The report also highlights the broader challenges facing AI companies as they push agentic capabilities that require deep OS-level integration. Unlike purely conversational or API-based interactions, computer use features depend on complex permission systems, accessibility APIs, and system-level hooks that vary significantly across operating system versions, hardware configurations (in this case, Apple Silicon M3), and security policies. macOS in particular has increasingly stringent sandboxing and privacy protections that can create edge cases—especially with newer OS versions or specific hardware—where accessibility permissions are technically granted but the underlying API calls that enumerate running or installed applications fail silently or return incomplete data. This class of bug is difficult for end users to self-diagnose, as the visible symptoms (permission dialogs, toggles) suggest everything is configured correctly even when the underlying system call is broken.
This kind of issue underscores a recurring theme in the rollout of agentic AI tools industry-wide: the gap between demoing a capability under controlled conditions and shipping it reliably across the long tail of real-world hardware, OS versions, and permission states. As Anthropic, OpenAI, and other labs race to make AI agents capable of operating computers autonomously, edge-case bugs like this one reveal that the "last mile" of OS integration—rather than the underlying model intelligence—is often the harder engineering problem. For Anthropic specifically, resolving such bugs quickly and transparently will be important for maintaining trust in computer use as a flagship differentiator, particularly as competitors race to ship similar desktop automation capabilities.
Read original article →