Bigarrow gives AI agents visual screen pointers on macOS

Developer Franz Enzenhofer has released bigarrow, an open-source command-line tool and agent skill designed to let autonomous AI agents point to on-screen elements on macOS. Built as a single MIT-licensed Swift binary requiring Xcode 16+ and macOS 14+, the tool provides native skill integrations for Claude Code and OpenAI Codex. The utility solves a persistent friction point in agentic workflows: when an agent encounters a step requiring human action—such as OAuth consents, system permission toggles, or two-factor authentication—it can render bright arrows, highlighted boxes, and custom text signs over target interface elements rather than printing text prompts to an unmonitored terminal.

The lightweight binary operates at the macOS screen-saver level, allowing drawn shapes to stay above active windows without stealing keyboard focus or capturing user keystrokes. Clicks pass directly through the overlays to underlying applications, and the elements automatically terminate based on configurable timers, process exits, or target clicks using flags like --until-click. Drawing requires zero system permissions, though identifying target coordinates or labels relies on Accessibility or Screen Recording permissions inherited from the agent's host terminal application. The overlay engine consumes roughly 1.4 percent CPU on continuous integration runners and includes built-in accessibility features such as text-to-speech notifications via a --say flag.

For developers and practitioners deploying agentic automation, bigarrow offers a lightweight visual interaction layer without bloating model context windows. The core skill definition consumes just 182 tokens, with full instructions adding 1,398 tokens as measured by Anthropic's Claude Opus 5.5 token-count API, alongside 1,008 tokens for optional visual properties. Verified across macOS versions 15, 26, and 27 with 104 automated tests and 18 behavior checks, the project gained nearly 400 GitHub stars within two days of launch. By providing reliable visual cues and copyable text blocks straight from agents to the screen, practitioners can comfortably run background automations knowing the agent can directly flag required manual interventions.