Pc Screen Control

A screenshot is a picture made for human eyes.This hands your AI the screen as data instead.

An MCP server for Windows · one-click install for Claude Desktop · works with any MCP client

---

Install

Needs Windows and Python 3.9+ — tick "Add python.exe to PATH" during its setup. Two downloads, the same server inside both. Take either.

pc-screen-control-setup.zip — works everywhere. Extract it, double-click INSTALL.bat, restart your AI. One run registers every MCP client it finds and prints a line per client saying added, updated or skipped: Claude Desktop, Claude Desktop from the Microsoft Store, Claude Code, and ChatGPT desktop / Codex. For Cursor, VS Code, Cline and Zed — which want a config block pasted in — print-config.py beside it prints that block with your real path already filled in.

pc-screen-control.mcpb — one click, Claude Desktop only. Settings → Extensions → Advanced → Install extension → restart Claude. Convenient for a first install. It cannot repair a broken one: if Claude answers "Private dir leaf redirects (junction/substitute-name plant)", then no version of this file will install — not the new one, not the old one you had before. The block is in Claude, not in the package. Use the .zip; that route is unaffected ([why](docs/WINDOWS-EXTENSIONS.md)).

Then ask your AI to run self_test. It reports what works, and every failure names its own fix.

Upgrading: install straight over the old version. Do not uninstall first — on Windows that is the step that leaves Claude unable to install the extension again. Quit the app completely, tray icon included, then install. · [Full guide](docs/GUIDE.md) · [Other clients](docs/OTHER_CLIENTS.md) · [If your antivirus objects](docs/ANTIVIRUS.md)

---

What it does

A screenshot is a picture. An AI looking at one has to work out what the shapes mean and then aim at a pixel.

Windows already publishes the same screen as structured data — every button, field and list, by name, type and state. It is what screen readers read. This hands your AI that instead of the picture.

And a button is not clicked at a coordinate. It is asked to press itself: the application carries out its own action and reports what changed. Nothing is aimed at, nothing is guessed, and the confirmation is in the reply rather than in another screenshot.

| | screenshot + coordinates | this | |---|---|---| | find a button | work out which shape it is, aim at a pixel | ask for it by name | | press it | move the mouse there and click | tell the application to invoke it | | the window moved | the coordinates are wrong | nothing changes — the name did not move | | did it work? | take another screenshot and look | the reply carries the state before and after | | your mouse | taken | untouched |

---

Does it use your screen?

Mostly not.

30 of the 34 tools never touch your mouse or keyboard. Reading the window tree is pure data. Pressing a button, filling a field, setting a slider go through the accessibility interface and work on a window that is behind others, or not visible at all.

Four tools do take the hardware: click, drag, send_keys, hold_key. They are for surfaces that paint themselves and publish no controls: editing canvases, timelines, games. While one runs, the screen edge pulses red, a Windows notification states what is happening and for how long, and your input is held so your keystrokes cannot land inside the work. Window, focus and text cursor are restored when the block ends, and the reply reports whether that succeeded.

A claimed window is outside your view entirely. claim_window moves a window just past the edge of every monitor. It keeps running and stays fully operable by name, but it is not visible and Windows will not let your pointer leave the monitors, so you cannot click it by accident. Work on a parked window takes no block at all: no pulse, no held input, no interruption.

This is not a private virtual display. Windows does not let an ordinary program create a second screen for an application. Parking off-monitor is the closest equivalent that works.

The tray icon's Pause, Stop and Watch stay clickable while your input is held; the taskbar is excluded from the hold.

---

The 34 tools

| | | |---|---| | describe_screen | Every window, classified readable / shallow / canvas-only. Start here. | | list_windows read_ui_tree find_elements | The control tree, searchable, each node with a ref | | element_from_point get_focus get_text read_text | What is where, what has focus, what it says | | read_table | A grid or details list as rows, columns and headers | | capture | Image of the screen, a window, or a single element | | invoke toggle expand select set_text | Operate controls — all return before/after | | set_value | A slider, spinner or scroll position to an exact number | | window | Move, resize, minimise, maximise — without the mouse | | clipboard | Read or write it. One call instead of hundreds of keystrokes | | menu | Open a context menu and read it — menus do not exist until opened | | click drag scroll | Coordinate input, last resort for self-drawn surfaces | | send_keys hold_key | Keyboard, for shortcuts | | wait_for wait | Wait for a condition, not for the clock | | batch | Several verified steps in one call | | launch_app close_window focus_window | Processes and windows | | self_test | Checks the installation and reports what is wrong and how to fix it | | claim_window release_window | Park a window out of your reach, and put it back to the pixel | | set_guard | Who has priority while the AI works — claude or me |

---

Reach, measured

Measured with tests/measure_desktop.py, which ships here so the numbers can be reproduced or contradicted on your own machine.

| | actionable nodes | |---|---:| | File Explorer | 220 | | Chrome / Electron — VS Code, Slack, Discord, Teams, Notion | 207–398 | | DaVinci Resolve, Project Manager | 53 |

Browsers and Electron build their tree only once something asks: a first look reports 13 nodes, 207 after waking it. Self-drawn surfaces publish no controls and are reached with capture and click instead.

---

Before you rely on it

• Windows only. docs/PORTING.md maps the patterns onto the macOS

Accessibility API, but none of it has been implemented. macOS is not available.

• No network at all. The server opens no socket — no update check, no

telemetry. Two files here do use the network and neither ships: the updater you start by hand, and a workflow on GitHub's machines that verifies every release against its source. [SECURITY.md](.github/SECURITY.md)

• Control names follow the window's language — Speichern, not Save.

About half of all elements also carry an automation_id, which does not translate; find_elements searches both and says which matched.

• Your antivirus may object. A program that reads other applications and

moves the mouse is structurally similar to spyware. There is no code signing certificate here; the server is a single readable Python file instead. [docs/ANTIVIRUS.md](docs/ANTIVIRUS.md)

• Administrator processes are invisible. Windows blocks input across

integrity levels by design.

• No undo of its own. Closing a window or sending a message is not something

any tool takes back.

• Maintained by one person. No support contract, no SLA.

---

Verification

Every claim on this page has a test behind it. 19 test files run in CI on Python 3.9, 3.11 and 3.13, and each release is checked against its own source automatically: the published packages are downloaded and compared byte for byte with the code at their tag.

Run them yourself: python -m pytest -q, or any single file directly — python tests/test_offline.py.

[Changelog](CHANGELOG.md) · [Security](.github/SECURITY.md) · [Contributing](.github/CONTRIBUTING.md) · [Other clients](docs/OTHER_CLIENTS.md) · MIT © 2026 Desteny Development

---

Disclaimer. This software controls your computer — buttons, typing, closing windows, dragging, including in apps holding unsaved work. Provided as is, no warranty ([LICENSE](LICENSE)). You are responsible for what you automate; test on something you can afford to lose. Automating third-party software may breach its terms.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: desteny-dev
• Source: desteny-dev/pc-screen-control
• License: MIT

Install and usage instructions live in the source repository linked above.