Computer Use

The Agent Sees
Your Screen

"Point it at any application by name. It clicks, types, drags, and reads back what changed."

Any Application
14 Interaction Tools
Full KB & Mouse

How It Works

Scan once. Act many times.

The monitor tiles your screen into regions, runs each through a vision model, and builds a persistent element map. That map stays live until the screen actually changes. Incremental updates mean a single chess move rescans 2 tiles, not the whole screen.

Step 01

Tile & Scan

The screen is divided into vision tiles. Each tile runs through a vision model and returns labelled elements with positions — Submit button, File menu, search input. The full map persists until something changes.

Step 02

Semantic Interactions

Every action uses natural language: clickElement("Save button"), dragElement("D2 pawn", "D4 square"), typeIntoElement("search box", "hello"). The agent never sees or passes pixel coordinates.

Step 03

Built-In Feedback

Every action returns whether the screen changed, by how much, and how long it took to settle. Background animations — CSS transitions, floating particles — are automatically subtracted so the agent knows if its click did something vs. hit dead space.

Step 04

Incremental Updates

After the initial scan, only tiles that visually change get re-scanned. A chess move rescans 2 tiles, not the whole screen. Large transitions like navigating to a new page automatically invalidate the full cache and trigger a fresh scan.


The Interaction Toolkit

Full keyboard and mouse coverage, plus timing primitives for flows that depend on UI settling before the next action.

Click
Left-click any element by name — auto-resolved to screen coordinates, no pixel math
Double-Click
Double-click for open, select-all, and rename actions — same semantic targeting as click
Right-Click
Context menus, inspect options, custom actions — right-click any labelled element by name
Drag
Smooth 10-step interpolated drag from element to element — chess, sliders, file drag-and-drop, reorder lists
Hover
Trigger hover states, tooltips, dropdown previews without clicking
Scroll
Scroll at element position — positive scrolls up, negative scrolls down, configurable notch distance
Type
Find input by name, click to focus, then type. Long text auto-pastes via clipboard for speed. SendKeys specials supported.
Key Combos
Full keyboard shortcuts — Ctrl+S, Alt+F4, Shift+Tab, function keys, any combination
Hold Key
Hold a key for a specified duration — shift, ctrl, alt, or any key the app needs held
Clipboard
Read what is in the clipboard or write to it — confirm what was copied, pre-load text for paste
Wait-For-Element
Poll until a named element appears on screen — dialog buttons, loading spinners, status text
Wait-For-Change
Wait for any visible screen change after an action — configurable sensitivity strips ambient animation noise
Wait-For-Gone
Wait for an element to disappear — verify a dialog closed, a spinner dismissed, a banner faded
Mouse Position
Read the current physical cursor position — useful for diagnosing coordinate mismatches between element map and actual click target

Compatibility

If you can see it, the agent can interact with it.

Computer Use is not a browser automation tool or a web scraper. It operates at the OS input level using native Windows APIs, so it works with anything rendered on screen — not just web content.

Desktop Apps
Native Win32, WPF, Qt, Electron — any windowed application. Menus, dialogs, toolbars, sidebars. If it renders pixels, the agent can read and click it.
FULLY SUPPORTED
Web Apps & Browsers
Any browser, any page. Works on top of any site without needing extensions or browser access. Vision sees the rendered page — dynamic content, SPAs, canvas — all included.
FULLY SUPPORTED
Games & Terminals
Chess boards, strategy games, command-line interfaces, SSH sessions in terminal emulators. The drag tool handles piece moves; type handles commands. Not limited to UI-framework apps.
SUPPORTED

Honest Context

What it is — and what it is not.

Computer Use is observe-think-act, not continuous control. It is built for automation workflows, not reflex-driven interaction. Here is what to expect.

Speed & Resolution

A 2560×1440 screen produces 8 vision tiles per full scan. Cloud models process tiles in parallel — 2–3 seconds. Local vision models run sequentially — expect 20–40+ seconds depending on hardware.

Element Recognition

Label quality scales with the vision model. Labels can be inconsistent between scans — “Submit” vs “Submit button” — and small or visually ambiguous elements may be missed or mislabelled.

Windows Only

Uses native Windows input APIs. macOS and Linux are not supported in the current release.

Single Monitor per Session

Multi-monitor setups work, but you start a monitor on one display at a time. You can switch between monitors; simultaneous dual-monitor sessions are not currently supported.

Not Real-Time

Each interaction round takes seconds, not milliseconds. Great for automation workflows and UI testing. Not suitable for twitch gaming, real-time video playback control, or anything requiring sub-second reaction.

Observe-Think-Act

The agent scans, reasons, then acts — one step at a time. It does not maintain a continuous stream of screen awareness between actions. Each action is a deliberate, complete round-trip.

What this is not: This is not a macro recorder, a script runner, or a remote-control tool. It is an AI agent that observes, reasons about, and interacts with your screen — best used for automating repetitive multi-step UI workflows, testing interfaces, or any task where the sequence of actions depends on what the screen shows at each step.
"If you can see it, the agent can interact with it."
— that is the entire design goal

Automate anything on your screen

Any app. Any workflow. Zero coordinates.

Buy Now — $79 ← Explore More Features