Computer Control

Let agents operate native desktop applications through an accessibility-first driver

Overview

Superset includes a computer skill that teaches agents to operate native desktop applications through an accessibility-first driver. It can inspect windows, read accessibility state and screenshots, click, type, scroll, use native menus, and verify the resulting UI state on macOS, Windows, and Linux.

The skill's workflow is driver-agnostic; Cua Driver is the default. On macOS, Peekaboo works as an alternative — agents use whichever supported driver you have installed, and ask before installing one.

The agent works in the actual app window you can see. Cua Driver targets a specific process and window and uses background interaction when the platform supports it, so routine automation does not need to take over your mouse or bring the target app to the front.

Install a driver

Optional third-party software

Drivers are not built or maintained by Superset — Cua Driver comes from Cua and Peekaboo from its open-source project — and nothing in Superset requires one: without a driver installed, this skill simply isn't used. Installing a driver and granting it Accessibility and Screen Recording gives it, and agents driving it, deep control of your machine — extend it the same trust as any system-level tool.

The skill is bundled with Superset, but the driver runtime is installed separately. Cua Driver is the default:

macOS and Linux

/bin/bash -c "$(curl -fsSL https://cua.ai/driver/install.sh)"

Windows (PowerShell)

irm https://cua.ai/driver/install.ps1 | iex
cua-driver autostart enable
cua-driver autostart kick

Then verify the host:

cua-driver --version
cua-driver doctor

macOS also requires Accessibility and Screen Recording permission for the CuaDriver app. Run cua-driver permissions grant and follow the System Settings prompts. See the complete installation and permissions guide.

Cua Driver enables anonymous telemetry by default; turn it off with cua-driver telemetry disable.

Prefer Peekaboo? (macOS only)

brew install steipete/tap/peekaboo

Peekaboo needs the same Accessibility and Screen Recording permissions. Agents use whichever supported driver is installed, so installing Peekaboo instead of Cua Driver selects it.

Hand an app to your agent

Ask naturally, or invoke /superset:computer in Claude Code and /superset/computer in Superset's in-app chat:

Open Calculator, compute 27 × 14, and tell me the displayed result.

Use the open design app to export the current canvas as a PNG, then verify the
file appeared in the chosen folder.

The skill makes the agent declare a task-scoped session, identify the exact app and window, take a fresh state snapshot before every action, and verify the UI afterward. It starts with accessibility targets, uses pixels only when needed, and escalates to foreground or full-desktop control only after observed failure.

Computer or browser?

Use browser for a page open in Superset's in-app browser pane. It provides direct DOM, console, screenshot, and Chrome DevTools Protocol access.

Use computer for native applications, the Superset desktop UI itself, browser chrome, or a browser engine that is not available through the in-app browser skill.

Safety

The agent sees the real state of the targeted application. The skill tells it to avoid unrelated private data, preserve your other windows, and ask before consequential actions such as sending a message, submitting a form that creates an external commitment, publishing, purchasing, deleting data, or changing account settings.