Cua routes agent clicks through OS-native events to leave the developer's mouse where it is

PyAutoGUI and its relatives synthesize physical hardware events, which the operating system then routes the way it routes a real click: to whatever window has focus, with the cursor moving there. The post describes the result plainly. Windows pop to the foreground, and you cannot type without interrupting the agent ^[5].

Background delivery is a different contract. The event goes to a named window or application, the pointer stays where you left it, and the focus of your editor is not taken ^[6]. Cua Driver claims that across macOS, Windows and Linux with one qualifier attached: "where supported by the platform" ^[6]. The post does not enumerate which of the three platforms qualify for which action, and it includes no benchmark results ^[15]. Clicking a button, typing a value and reading a native window's contents are three different API surfaces on each OS. Verify them separately on whatever you actually target.

The documented install is a remote script piped into a shell: `/bin/bash -c "$(curl -fsSL https://cua.ai/driver/install.sh)"`, or `irm https://cua.ai/driver/install.ps1 | iex` on PowerShell ^[7]. An input driver needs privilege on all three platforms, so that one line is where the trust boundary sits. The post also suggests that users of Claude Code, Codex, Cursor or Antigravity instruct the agent to install and integrate Cua automatically ^[8].

Where the agent runs is the more consequential choice. The post argues that the local physical workstation is undesirable for untrusted execution or for scalable evaluation runs ^[10]. Fleets answers that with claimed sandboxed Linux cloud desktops drawn from a pre-warmed pool. Lume answers the macOS half by managing local VMs on Apple Silicon through Apple's Virtualization.Framework, booting clean macOS Tahoe images from Apple restore images ^[9][11].

CUA-S1 is the cost argument. Per the post, using a general model token by token to decide which form input to click is slow and expensive. CUA-S1-FORMS scores UI element decisions directly from structured accessibility trees and document elements at sub-second speed ^[12][13]. The stated input is the tree. For that figure to transfer to your workload, the application under test has to publish an accessibility tree the model can read. An app that draws its own widgets gives it nothing to score, so the decision falls back to the general model the post calls wasteful ^[13][12].

Cua Bench is the piece that makes any of this measurable: define desktop tasks, run reference solutions, verify completion with evaluators, and export step-by-step trajectories to train future models ^[14]. Across the five modules, the only performance descriptor on offer is the word "sub-second" ^[16].