screen_view
The live screen, one window or a region, as a real image with coordinate rulers. An unchanged frame isn’t sent again.
desk-pilot
Claude Code plugin · Windows onlydesk-pilot sees the actual screen, reads each window’s controls and drives the real mouse and keyboard. It does it with three tools, on purpose: every tool schema costs tokens.
01
Run the downloaded file and follow the wizard.
claude) and Node.js 22.19 or newer are installed.claude plugin marketplace add acus9345/desk-pilot and claude plugin install desk-pilot@desk-pilot for you.claude plugin marketplace add acus9345/desk-pilot
claude plugin install desk-pilot@desk-pilot
/plugin marketplace add acus9345/desk-pilot
/plugin install desk-pilot@desk-pilot
02
From the project’s README.
screen_viewThe live screen, one window or a region, as a real image with coordinate rulers. An unchanged frame isn’t sent again.
desktop_controlClick, drag, scroll, type, keys, focus, wait, move or resize windows, and read a window’s UI Automation controls.
desktop_sequenceMany actions in one call, checked step by step, with one frame at the end.
Click a control by name or index. It is looked up again right before the click, and a disabled or stale one is refused.
Every input reports where it actually went, type reads the text back, and expect makes a step fail if the effect never shows.
Risky batches need your approval, a banner shows while Claude acts, and Esc stops it between steps.
In the README’s measurement (2026-10-25), desk-pilot added +191 tokens per session with tool search on, against +327 for Windows-MCP. Full comparison ↗
03