Image: GitHub
Why it mattersAn agent that reaches a login dialog or a 2FA prompt stops being a black box in a terminal and starts pointing at the exact window and button the human needs to touch next.
An AI agent that works for twenty minutes on your Mac and then prints "please click Allow" into a terminal you are not looking at wastes both of you. Franz Enzenhofer's bigarrow, a one-day-old Mac command line tool, lets a coding agent draw a click-through arrow with a sign at the exact button it needs you to press.
The repository is franzenzenhofer/big-arrow-on-the-screen on GitHub. The first commit is a day old, the project is MIT, and the repository held 58 stars when this item was written. The Show HN thread passed 65 points in its first hour.
Three verbs and a hard rule
The command set is three verbs. One points at a button by its Accessibility label and the app that owns it. One points at a screen coordinate. One draws an arrow that stays until the agent calls stop. Every arrow has a hard time limit, eight seconds by default and three hundred for the long form, and every arrow also goes away when the agent process that drew it exits. An agent that forgets to clean up cannot leave arrows behind.
Drawing itself needs no macOS permission. bigarrow runs as a small Swift binary with no daemon, no menu bar icon, and no telemetry. The README states that twice, and the author's follow-up: "it is an arrow". Reading window titles, finding a button by label, and listening for a click on the target use Accessibility, which macOS grants to the terminal the shell runs in, not to bigarrow. The author added bigarrow doctor to say which app holds that permission and which does not.
The gap it fills
Three situations the author names, and all three are familiar to anyone who runs Claude Code or Codex for real work. A permission dialog or OAuth consent screen the agent must not press. A two-factor code or payment confirmation the agent is not supposed to answer. One of fourteen Chrome windows that the agent knows is the right one, but the human cannot find. The CLI accepts a Chrome window title, raises that window first, and points the arrow inside the right place. A spoken option reads the sign aloud through the Mac's speech system, for the case where the human is making coffee.
Everything the arrow needs is in the invocation. There is no agent framework and no SDK. The CLI returns exit code 0 on success, 2 on bad input, 3 if the target cannot be found, and 4 if a permission is missing, so an agent that gets a 4 back knows which permission to ask the human for.
The project ships a Skill for Claude Code and Codex that tells the agent when to point, how to pick a target, to write a full sentence on the sign, to speak the sign when the human is probably not looking, and to clear the arrow once the human has acted. The author reports 87 automated tests, 17 behaviour checks on a clean runner, and a transcript where a fresh agent found the Chrome Reload button from the Skill alone and built the right command without help. The CPU cost of a pulsing arrow was 1.4 percent on a continuous integration runner.
A team running coding agents on a Mac gets a signal a desktop notification cannot carry: the exact window and the exact button the agent is waiting on.
Source
The repository is at franzenzenhofer/big-arrow-on-the-screen under the MIT licence.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.