OS Automation MCP servers

Inspect, debug and drive running WPF (.NET desktop) apps: visual tree, dependency properties, data bindings and binding errors, DataContext, and screenshots with open popups/menus included, plus clicking, item selection, text input and keyboard shortcuts. Auto-injects into any running WPF process (x64/x86) with no source changes. `dotnet tool install -g WpfVisualTreeMcp`

Gives a coding agent a build → see → interact → debug loop over the native GUI app it is writing: launch the app, screenshot it, read its accessibility tree, inject mouse/keyboard/gestures, tail its logs, and diff frames to confirm a change landed. Drives apps as an external black box with no app integration, so any toolkit or language works. 30 tools with text-first responses (element ids, diffs, waits) so the agent need not screenshot every step, plus optional sandboxing of the launched app. Backends: X11, Wayland, Windows, macOS, Android emulators, and the iOS Simulator.

Let an agent use a real iPhone, iPad or Mac through Apple Shortcuts: run a Shortcut on a locked phone and get the result back, send lock-screen notifications, and read what the phone sends (Calendar, Reminders, Notes, HomeKit, Wallet events). End-to-end encrypted relay, one automation on the phone. `npx -y askew-mcp`

MCP server for [FreezeText](https://store.juergenkoller.software/en/apps/freezetext) — OCR anything on your Mac screen. Freeze the screen and extract text via Apple Vision (videos, popups, protected PDFs), OCR a region or base64 image, and manage a searchable capture history. 12 tools.

Hybrid computer-use runtime. Fuses accessibility tree + Chrome DevTools Protocol + vision into structured context with per-element confidence. 4 MCP tools (see/act/think/perceive). Continuous awareness engine (Cortex) with freshness + side-effect detection. Works offline with Ollama + local models.

GNOME desktop automation for AI agents. 30 tools via D-Bus: screenshots, window management, mouse/keyboard injection, clipboard, workspaces, and system notifications. Works on any GNOME 45–49 Linux desktop.

Windows desktop automation for LLM agents with entity-based actions instead of coordinate-only clicking. Uses UIA, CDP, screenshots, keyboard/mouse/clipboard, and terminal control, plus entity leases, verified delivery, causal context, and interaction memory to reduce silent UI automation failures.

Local MCP stdio server that gives trusted AI clients read-only desktop context through OpenWand, including selection, clipboard, active-window, browser-page, and screen context.

Accessibility-first Computer Use for any agent: element ids instead of guessed pixels, background input that never moves your mouse, an agent pointer overlay, zoom with correct coordinate mapping, and an own tab group in your signed-in Chrome. Same 30 tools on macOS, Windows and Linux; npm launcher `munim-computer-use`.

Physical USB HID keyboard/mouse control via a Raspberry Pi Pico 2 running open-source firmware. Exposes move, click, drag, type, key combos, and scroll as MCP tools for any MCP client. Genuine physical HID input on the standard driver path, with a `--mock` mode for hardware-free trials. `pip install clawtouch-mcp`

MCP bridge for [Nemeton](https://store.juergenkoller.software/en/apps/nemeton) — native macOS virtual machine manager built on Apples Virtualization.framework. Create/control Linux & macOS VMs (no Parallels, no QEMU), CoW snapshots on APFS, 50+ MCP tools across VM lifecycle, console, files, networking, and host metrics.

Fork, break, and revert real VirtualBox VMs as MCP tools. Forks are linked clones (~0.1s) and revert restores a RAM snapshot in seconds. Drives GUI-only guests through the hypervisor with screenshot, click, scroll, type and keys, so an agent controls a machine with no SSH, guest agent, or network, including Windows XP. Windows host with VirtualBox required. `npx -y terrarium-mcp mcp`

Turn a Mac into a remote computer-use MCP server with one command. `uvx usemymac up` wraps CUA's computer-server (screenshot, mouse, keyboard, windows, apps, accessibility tree) in a per-run token, hides shell and file tools unless opted in, and opens a Cloudflare quick tunnel so Claude Code, Codex, or any MCP client on another machine can use the Mac. Rides an existing CuaDriver daemon when one is running.

Safe-by-default control of a Mac: read files, list processes & apps, take screenshots, read the clipboard; and behind explicit opt-ins, run shell commands (allowlisted), AppleScript, move files to Trash, and inject GUI input. Read-only by default with layered access modes, per-path/command allowlists, and human confirmation (MCP elicitation) on the dangerous calls. `npx -y @dockndevai/mcp-macos`

Full control of a Mac — drive it like a human: shell, AppleScript, files, processes, and human-like GUI (move, click, double/right-click, drag, scroll, type, key combos) with a screenshot perception loop. Full control by default; MACCTL_SAFE_MODE=true inverts to safe-by-default (read-only, gated, confirmations). The powerful sibling of mcp-macos. `npx -y @dockndevai/mcp-mac-control`