Computer Use Turbo

What Computer Use Turbo MCP server does

Computer Use Turbo MCP server connects an MCP client to desktop applications on macOS, Windows, and Linux. It can discover available applications, inspect a window, interact with controls, enter or paste content, use keyboard shortcuts, operate menus, and wait for visible or structural changes.

The agent receives an app representation as numbered accessibility-tree lines together with a screenshot. This lets it target a reported element or coordinate instead of relying only on image interpretation. The server can also expose system panels such as Control Center, Notification Center, and Spotlight through available menu-bar status items.

How it works

The Node.js MCP server communicates with the agent over standard input and output. It validates tool arguments, identifies the connected agent, starts the helper when needed, and converts screenshots into MCP image content. A platform-specific helper performs the actual application access, screen capture, input synthesis, and safety enforcement.

On macOS, the helper uses Accessibility APIs, ScreenCaptureKit, and CGEvent. Windows uses UI Automation, PrintWindow, and SendInput. Linux uses AT-SPI, with XTest on X11 and desktop portals, PipeWire, and libei on Wayland.

Applications can remain behind other windows during most operations, and the agent does not move the user's mouse. A live preview can show the activity. Certain operations, including some menu-bar interactions, file drags, and applications with custom-rendered interfaces, may briefly bring an app forward while the user is idle.

Setup and configuration

Installation requires Node.js 20 or newer and Git. Clone the repository, build and install the helper for the host operating system, then add server/src/server.mjs as an MCP server using the name turbo.

macOS requires version 14 or later plus Xcode or the Command Line Tools. Accessibility and Screen & System Audio Recording permissions must be granted in System Settings. Windows requires Rust with the MSVC toolchain and does not require permission grants. Linux requires Rust, an accessibility-enabled desktop session, and development packages for PipeWire, xkbcommon, and clang. Wayland also requires a desktop portal with RemoteDesktop and ScreenCast support.

Claude Code and Codex provide documented commands for registering the server. Other MCP clients can use the generated configuration from scripts/mcp-config.sh or its PowerShell equivalent on Windows.

Tools and capabilities

Computer Use Turbo MCP server exposes tools for the full interaction loop:

• Discover applications and inspect their current interface with find_apps and observe_app.
• Click controls, type or replace values, select field text, scroll, drag, and paste plain text, Markdown, or HTML.
• Send key chords, invoke named accessibility actions, and search or run application menu commands.
• Wait for text, element changes, windows, or application settling before continuing.
• Batch checked actions with run_steps, display or hide the preview, inspect permissions, and end a session with finish.

Later observations report changes rather than repeating the entire state where possible. run_command can execute a menu command without opening the menu, while wait_for responds to application events instead of requiring polling.

Limitations and notes

Password fields are not typed into automatically. Password-manager use requires confirmation, and system authentication remains protected. The user's agent permissions remain in control, and the helper is the only component that directly touches applications.

A signing certificate is recommended for macOS because rebuilding without one can require permissions to be granted again. Linux Wayland setup depends on a compatible desktop portal. Applications that do not expose usable accessibility information or that draw their own interface may require foreground handling and may provide less reliable targeting.