codex-computer-run

name	codex-computer-run
description	Use this skill when Codex needs to operate or inspect a signed-in desktop through the Codex Computer Run MCP server, including screenshots, visible window discovery, cursor position checks, mouse movement, clicking, scrolling, keyboard shortcuts, single-key presses, or Unicode text paste into focused applications.

Codex Computer Run

Overview

Use the Codex Computer Run MCP server as an observation-first control layer for a real desktop session. Its tools affect the active machine, so inspect context before acting and verify after actions that can change state.

Tool Discovery

• Prefer the mcp__codex_computer_run__ namespace when available.
• If tools are deferred, search for ComputerRun, codex computer run, or desktop screenshot mouse keyboard and choose the namespace that exposes the complete tool set.
• Expect these tools: screenshot, list_windows, cursor_position, move_mouse, click, scroll, press_key, hotkey, and type_text.
• If the MCP tools are unavailable, state that the Computer Run server is not configured in the current session instead of simulating desktop interaction with unrelated shell commands.

Platform Notes

• Windows is the native implementation and uses Win32 desktop APIs.
• Linux and macOS use best-effort command adapters. If a tool reports a missing dependency such as xdotool, gnome-screenshot, cliclick, or osascript, explain the dependency instead of retrying unrelated commands.
• The server must run in the signed-in graphical session for the desktop it controls.

Operating Protocol

1. Observe before acting:
   • Use list_windows to identify visible applications and likely targets.
   • Use screenshot when visual layout, coordinates, or UI state matters.
   • Use cursor_position before relying on the current pointer location.
2. Plan in absolute desktop coordinates:
   • Treat coordinates as desktop coordinates, not browser or app-relative coordinates.
   • Read screenshot metadata for platform, left, top, width, and height; multi-monitor layouts can have negative left or top values.
   • Move or click only when the target application and coordinates are known.
3. Act with the narrowest tool:
   • Use move_mouse for hover or to position before a click.
   • Use click for buttons, menus, tabs, selections, and context menus.
   • Use scroll for pages, lists, combo boxes, and scrollable panes.
   • Use press_key for one key such as enter, tab, escape, f5, arrows, or a single character.
   • Use hotkey for shortcuts such as ctrl+l, ctrl+shift+p, alt+tab, or ctrl+shift+escape.
   • Use type_text for text entry; it pastes Unicode through the clipboard and is faster and more reliable than repeated key presses.
4. Verify after meaningful actions:
   • Use screenshot after navigation, clicks, scrolls, or text entry when the resulting state matters.
   • Use list_windows again after task switching or launching apps.

Screenshots

• Use screenshot with include_image: true when the model needs to inspect the pixels.
• Use include_image: false with a path when only dimensions, a saved artifact, or later local inspection is needed.
• Save screenshots to an explicit temporary or workspace path when they may be referenced in the final answer.

Safety Rules

• Remember that mouse, keyboard, and clipboard actions affect the user's signed-in desktop.
• Do not perform destructive UI actions, submit forms, send messages, make purchases, delete files, or change account/security settings unless the user explicitly asked for that exact outcome.
• Confirm the intended foreground app with list_windows or screenshot before typing or pressing shortcuts that could affect the wrong application.
• Treat type_text as clipboard-changing; use it only when pasting into the focused target is intended.
• Keep delays short but use the optional delay parameter after actions that trigger UI transitions.

Quick Checks

For a non-destructive connectivity test:

1. Call list_windows with a small limit and confirm visible window metadata is returned.
2. Call cursor_position and confirm JSON coordinates are returned.
3. Call screenshot with a saved path and include_image: false; confirm screenshot metadata includes the saved path and virtual desktop bounds.
4. Optionally move the cursor by a small, reversible amount with move_mouse, then call cursor_position again to confirm the new coordinates.