Guides

Guides on Windows desktop automation, accessibility-tree control, UI automation, and AI agents that drive real apps.

• A TV to use as a computer monitor: the developer-side answer the RTINGS-style guides never write

  Every top result picks a TV by input lag and chroma 4:4:4. The piece they skip: a TV is just one endpoint of the display spectrum a desktop automation framework has to handle. The other endpoint is no monitor at all. Terminator

• A UI automation test you can write as a boolean expression: inside Terminator

  Most UI automation test runners treat a locator as a flat string: one id, one role, one XPath. Terminator parses selectors as a boolean expression. role:Button && (name:Confirm || name:Best\u00e4tigen) && !visible:false is tokenized, shunted through a three-precedence Shunting Yard parser (OR=1, AND=2, NOT=3), flattened, and evaluated as a single predicate on each element. Source: selector.rs, 753 lines.

• A Windows automation script you record instead of write: 14 semantic event types, replayable against the accessibility tree

  Every other guide on this topic tells you to type a PowerShell or AutoHotkey script. Terminator

• Accessibility API desktop automation: fire Control Patterns, skip the mouse

  Most guides treat the accessibility tree as a read-only inspection surface. The unlock is that UIA also ships Control Patterns (Invoke, Toggle, ExpandCollapse, Value) that act on elements without moving the mouse. Terminator

• Accessibility API for AI agents: stop re-reading the whole tree, diff it

  Most articles about driving an AI agent through the accessibility API stop at

• Accessibility API for computer use agents: the seven-mode click_element router

  A computer use agent that only clicks by accessibility selector breaks the moment it hits an Office canvas, an Electron child window, or a custom-rendered control. Terminator

• AI agent computer use reliability: push retries into the tool schema, not into the prompt

  Most computer-use stacks expose action verbs only (click, type, screenshot). The model has to write retry and verification logic in its own chain of thought, every call. Terminator

• Automation desktop application with five detection sources fused into one indexed tree

  Every other automation desktop application tool commits to one view of the screen: UIA, or DOM, or computer vision. Terminator does not. Its ElementSource enum has five variants (Uia, Dom, Ocr, Omniparser, Gemini) with the prefixes u, d, o, p, g, and cluster_elements runs union-find to merge bounding boxes that sit within 1.5x the smaller element

• Automation for desktop application: the Boolean and spatial selector language Terminator compiles through a Shunting Yard parser

  Most automation for desktop application work still binds to coordinates or image matches. Terminator ships a typed selector language with Boolean algebra (&&, ||, !) and spatial relators (rightof, leftof, above, below, near) that a Shunting Yard parser compiles into an AST and runs against the OS accessibility tree. Source: crates/terminator/src/selector.rs lines 205-272; near threshold at crates/terminator/src/platforms/windows/engine.rs line 1815.

• Automation for Windows that your AI coding assistant can actually drive: 35 MCP tools over UI Automation

  Most automation for Windows is a canvas for humans or a hotkey language for scripts. Terminator is a developer framework that exposes Windows as 35 typed MCP tools to Claude Code, Cursor, and Windsurf. Every action returns a before/after UI tree diff so the agent can verify its own work.

• Automation in Windows that does not steal your caret: how Terminator saves and restores focus around every action

  Most automation in Windows fails the moment you keep typing. Terminator asks the OS for your focused element and caret range via UIA TextPattern2 before every action, then restores both when the automation finishes. Source: crates/terminator/src/platforms/windows/input.rs.

• Automation on Windows is slow because of IPC: how Terminator collapses a 6.5s tree walk to 200ms

  Every naive automation on Windows pays a COM round trip per property per element. Terminator

• Automation scripts for Windows: one API for every app on the desktop

  PowerShell and AutoHotkey cover sysadmin and hotkeys. They do not address Win32, UWP, WPF, WinForms, and Electron apps with one script. Terminator

• Automation testing tool for desktop application: a transpiler that returns AI-fixable JSON, not stack traces

  Most desktop test runners hand you a stack trace when a test fails to compile. Terminator hands the failing test back to whatever AI wrote it as a JSON object with error_type, error_line, fix, and a recovery action. Four recovery variants. Fourteen TypeScript-detection regexes. A condition parser that strips smart quotes before evaluating. The pieces of a desktop testing tool designed to be driven by an LLM that may have written the test five seconds ago.

• Automation tools for UI testing that prove they can see the UI first

  Every ranked list of automation tools for UI testing scores vendors on selectors, self-healing, and AI authoring. None of them ask whether the worker about to run the test can actually see a UI right now. Terminator

• Automation tools for Windows: the one that refuses to run an ambiguous selector

  Most automation tools for Windows happily search the whole desktop tree and match the wrong button in the wrong app. Terminator

• Automation tools for Windows: why the cache-batched accessibility tree beats screenshots, coordinates, and RPA studios

  Most automation tools for Windows either replay pixel clicks or drive a drag-and-drop studio. Terminator is a developer framework that builds the Microsoft UI Automation tree in one CacheRequest round-trip (ControlType, Name, BoundingRectangle, IsEnabled, IsKeyboardFocusable, HasKeyboardFocus, AutomationId), then serves every subsequent read from local cache with zero further IPC. That is the speed budget an AI coding assistant needs to drive a real desktop in a loop.

• Automation UI testing that does not stop at the browser: Playwright-shaped tests for every desktop app

  Most guides on

• Automation UI testing tools: the tree-diff primitive every 2026 roundup misses

  Every ranking of automation ui testing tools scores vendors on selectors, AI self-healing, and flaky-test retry loops. None of them cover the verification primitive that makes a test loop actually deterministic: diff the whole accessibility tree before and after each action, with volatile fields stripped so reruns produce identical output. Terminator ships this as simple_ui_tree_diff in crates/terminator/src/ui_tree_diff.rs. Here is what it does and why no one else has it.

• AXUIElement::system_wide() in the accessibility crate: the three patterns the docs don

  AXUIElement::system_wide() in the Rust accessibility crate is a one-line wrapper around AXUIElementCreateSystemWide. The hard part is the three patterns it does not give you: an unsafe Send + Sync wrapper, a permission gate via AXIsProcessTrustedWithOptions, and a two-step walk through AXFocusedApplication then AXFocusedUIElement. Notes from a 4,368-line macOS adapter that was deleted on 2025-12-16 and is still the most complete reference for this code path.

• Browser agents leaving the DOM: where the tool surface ends, and what to do about it

  Every browser agent breaks the moment its workflow crosses out of the DOM. The download triggers a Save dialog, the link opens a native app, the OAuth code arrives in a desktop authenticator. Terminator

• Browser automation hits a desktop ceiling: the seven moments Playwright goes silent, mapped to the OS-level tool that takes over

  Browser automation works until the next click is owned by a different process. Then it stops, hard. This page enumerates the seven concrete ceiling moments (save dialogs, system print, native authenticators, OAuth desktop redirects, fullscreen surfaces, Open With handlers, app-to-app handoff) and maps each one to the specific Terminator MCP tool that resolves it. Of 38 #[tool(...)] declarations in crates/terminator-mcp-agent/src/server.rs, exactly one runs inside the DOM. The other 37 exist because the ceiling exists.

• Can I use a TV as a computer monitor? Yes, and here is the automation-side story nobody on the SERP tells

  Every top result for this query is about HDMI cables, chroma subsampling, and input lag. The piece they skip: once a TV becomes monitor 2, any script or AI agent driving your desktop has to become monitor-aware. Terminator exposes a 9-field Monitor struct in crates/terminator/src/lib.rs:274 and a get_monitor_by_name() call so you can target a TV by its reported display name.

• Can I use a TV for a computer monitor? Yes, and here is what happens to every overlay your PC already draws

  HDMI cables and input lag are the easy part. The part nobody writes about: plugging a TV into your PC resizes the Windows virtual screen bounding box, and every piece of software that draws a full-screen overlay (including Terminator

• Can you use a computer monitor as a TV? Yes, and if it is plugged into a PC you can skip every adapter the other guides recommend.

  Every top-ranked guide sends you to a streaming stick, a set-top box, or a coax-to-HDMI tuner. If your monitor is already connected to a computer, your computer is the tuner. Terminator ships a 151-line example that automates VLC to open any YouTube livestream and hit Play, driven by an AI agent. No Fire Stick, no Chromecast, no remote.

• Claude computer use: the pixel-coordinate loop, and the selector-based alternative nobody explains

  Claude

• Coded UI automation in 2026: replacing Microsoft

  Microsoft deprecated Coded UI Test in Visual Studio 2019 and shipped nothing to replace it. Terminator brings a Playwright-style

• Computer use agent state tracking: why a 118-pixel window shift returns zero diff

  Most computer use agents track UI state by re-snapshotting the screen or tree after every action. Terminator captures the accessibility tree before and after, strips IDs and bounds with two regexes, then returns the structured diff as part of the action result. A window that moved 118 pixels but didn

• Computer use on native apps: what the accessibility tree actually looks like to the agent

  A computer use agent never sees the raw UIAutomation or AXUIElement tree of a native app. Terminator serializes it into a compact numbered list, #1 [Button] Save (bounds: [...], focused), where every clickable element gets a 1-based index. The agent emits the index, not pixel coordinates. Source: format_node in crates/terminator/src/tree_formatter.rs.

• Cross-platform desktop automation: how to verify the runtime is actually healthy on Windows and macOS before you click

  A scripted click that never lands is worse than a test that fails fast. This guide shows how Terminator

• Desktop application automation that reports whether the click actually did anything

  Most desktop application automation libraries expose click(element) as a fire-and-forget call. Terminator

• Desktop automation and the accessibility tree: what one node costs to capture

  The accessibility tree is not a free structure you read once. Each node property is a separate cross-process call. Terminator

• Desktop automation software: the recorder denylist that decides whether your workflow still runs in week two

  Most desktop automation software is graded on what it can record. Terminator

• Desktop automation tools: why one clustered view of UIA, DOM, OCR, Omniparser, and Gemini beats picking a single perception method

  Every desktop automation tool forces a choice: accessibility tree (WinAppDriver, UiPath) or pixel vision (SikuliX, Claude computer use, OpenAI Operator). Terminator fuses five perception sources (UIA, browser DOM, OCR, Omniparser icon detection, Gemini vision) into one spatially-clustered index so an AI coding assistant can target whichever source actually saw the element. This page walks through how that clustering works.

• Everyday computer-use modal failure mode: the OS already labels modals, vision agents don

  The single most common failure for pixel-loop computer-use agents (Claude

• Free desktop automation tools that let an AI assistant compile-check its own script before any click runs

  Most lists of free desktop automation tools rank products by license terms, feature checklists, and supported recorders. None of them ask the question that matters once an AI assistant is the one writing the script: can the tool tell the assistant the workflow does not even compile, before that workflow opens a window? Terminator ships an MCP tool, typecheck_workflow, whose entire job is to run tsc --noEmit on the assistant

• Free desktop automation tools, audited at the flag level: where the cost cliff actually hides

  Every free desktop automation tool has a cost cliff somewhere: seat caps, unattended-run paywalls, license servers, telemetry you cannot switch off, or a cloud backend that goes away if the vendor does. This guide walks through where that cliff sits for each of the usual products, then shows Terminator

• How to use two computer screens: the developer side of dual monitors nobody writes about

  Every dual-monitor guide explains cables and Display Settings. This one explains the other half: how software tracks which window is on which screen, why DPI scale factor changes per monitor, and the Windows UIA bug that silently broke clicks on screen two until April 2026. All grounded in Terminator

• Legacy desktop automation for SAP GUI and Oracle Forms without the RPA tax

  SAP GUI for Windows and Oracle Forms (via the Java Access Bridge) both publish Microsoft UI Automation trees. That means a UIA-based developer framework drives them with the same locator(role:..., name:...).click() you would use for Notepad. No SAP GUI Scripting COM, no Forms-specific connector, no per-bot license. Terminator

• LLM desktop automation in April 2026: every model release, plus the one-line replacement that decides whether their clicks actually land

  April 2026 brought a wave of computer-use LLMs (Microsoft Fara-7B, Gemma 4, Manus updates, Claude Mythos preview), but the consequential change for desktop reliability shipped on April 2 in Terminator 0.24.31 (PR #473). It replaced Windows UI Automation

• Locked Mac, computer use, stale screen state: what is actually frozen and what is not

  When the Mac locks mid-task, your computer use agent does not get a black screen, it gets a frozen one. The WindowServer stops compositing app windows to the screen buffer, so any screenshot returns the last rendered frame. The accessibility tree of background apps keeps responding normally, so the agent

• macOS accessibility UI tree automation: the write path nobody warns you about

  The read path is easy: AXUIElementCopyAttributeValue walks the tree. The write path is the trap. AXPress and AXClick return success on browser-rendered views and do nothing. A real AX automation engine ships an 8-browser bypass list, a 3-tier click fallback, and synthetic CGEvents. Notes from a 4,368-line macOS implementation that got deleted on 2025-12-16.

• MCP server list: where your client actually keeps one, and the 12-path map an installer has to know

  Every top result for

• MCP server means a process, not a protocol: the operational layer every explainer leaves out

  Every top result for

• Microsoft Power Automation Desktop, but the workflow file is plain YAML your AI assistant can write

  Power Automate Desktop locks your automation inside a proprietary designer. Terminator is a developer framework for the same job: the workflow is a YAML file with readable selectors like role:Button && name:Save, an MCP server Claude Code or Cursor talks to directly, and TERMINATOR_HEADLESS=true runs the same file unattended on a Windows VM.

• Microsoft UI Automation tool: how Terminator pre-fetches a whole subtree in one COM call

  Most Microsoft UI Automation tools die on per-property COM round-trips. Terminator builds an IUIAutomationCacheRequest with seven UIProperty fields, sets TreeScope::Subtree (Element=1, Children=2, Descendants=4 = 7), and reads everything via get_cached_*. Source: tree_builder.rs lines 398 to 466, MIT licensed.

• Notion MCP server, but for the desktop app: controlling Notion through the OS accessibility tree

  Every guide for

• Open source desktop automation projects, April 2026: which ones still ship, which ones speak MCP, and the only one with a real selector grammar

  A grounded look at the open source desktop automation projects that matter in April 2026. Most of the listicles re-list the same six tools from a decade ago. This guide groups projects by what they actually hand to an AI coding assistant: pixels, a Python wrapper, a flowchart studio, or a typed selector grammar with an MCP server. Terminator is the only one in the last category, and the implementation lives in 753 lines of Rust at crates/terminator/src/selector.rs.

• Playwright MCP server, then off the page: what an MCP server looks like when it does not stop at the browser tab

  Playwright MCP is excellent inside a tab. The moment your AI agent needs to leave the tab (open Excel, paste a value, run a CLI, drag a file into a desktop app) it is out of room. Terminator is the same shape of MCP server (accessibility-tree-based, click and type by selector, discovered at runtime), but its dispatch function spans the whole OS. And it ships a separate Chrome extension on ws://127.0.0.1:17373 so one workflow mixes DOM-level and native-app steps.

• Python automation on Windows: hand the whole window tree to an LLM in one call

  Every other guide on Python automation on Windows teaches imperative pywinauto loops. Terminator

• Python desktop automation that actually runs concurrently with asyncio.gather

  pyautogui, pywinauto, and the rest of the Python desktop automation libraries are all sync. Terminator

• Python for Windows automation: 16 typed exceptions instead of try/except Exception

  Most Python libraries for Windows desktop automation throw one or two exceptions and call it a day. Terminator

• Robotic desktop automation software graded on its selector engine, not its Studio screenshot

  Every roundup of robotic desktop automation software grades the same drag-and-drop Studios. Terminator is graded on what its bot actually uses to find a button on the screen: a 24-variant typed selector enum at crates/terminator/src/selector.rs that composes positional, boolean, and structural operators against the Windows UI Automation tree, exposed to AI coding assistants through 35 MCP tools.

• RPA accessibility tree selectors: the actual grammar, with operator precedence

  Most RPA selectors are XML blobs that pin to a specific window path. Terminator

• Selenium UI automation, stretched to every native app: how Terminator reuses your locator grammar outside the browser

  Selenium taught a generation of engineers to pick UI elements by role, name, and CSS selector. Terminator ships the same prefix grammar, the same >> descendant chaining, and five spatial operators (rightof:, leftof:, above:, below:, near:) that Selenium has no equivalent for, running against the OS accessibility tree in 753 lines of Rust.

• Sentence to desktop automation script: the structural reason it works in Terminator and fails when you ask a cold LLM

  A sentence becomes a runnable script in Terminator because the MCP server ships a system prompt with one load-bearing rule: never guess or predict element attributes. The AI must call get_window_tree first and resolve selectors against the actual tree. The script is a sequence of MCP tool calls expressible as YAML. Source: terminator-mcp-agent/src/prompt.rs line 21.

• Slack MCP server without a bot token: what driving the Slack desktop app through accessibility APIs looks like

  Every Slack MCP server you have seen wraps the Slack Web API. That means a bot token, a workspace admin install, a finite set of API scopes, and rate limits. Terminator is a different shape of Slack MCP server: it drives the running Slack desktop client directly through OS accessibility APIs, as the logged-in user. No bot, no OAuth, no scope list. This page explains how that works, what it unlocks, and where it stops.

• Task automation for Windows with spatial selectors: click the button right of the Email field, no coordinates

  Every Windows task automation tool except Terminator forces you to pick between reading the accessibility tree by role and name, or measuring pixel coordinates. Terminator

• Task automation on Windows: the four fields that turn a script into a resumable workflow

  Task Scheduler runs a .exe and reports pass or fail. Terminator

• Terminator + Claude: the two install paths nobody distinguishes (Claude Code vs Claude Desktop)

  There are two ways to wire Terminator into Claude: claude mcp add for Claude Code, and editing claude_desktop_config.json for Claude Desktop. Most guides only show the first one. Sourced from crates/terminator-mcp-agent/config.js: 12 supported MCP clients with exact config paths, with Claude as the default fallback.

• Terminator MCP: the desktop automation server that type-checks your workflow before it ever clicks a pixel

  Every other walkthrough of this stops at claude mcp add terminator and a tool list. None of them tell you that Terminator

• Terminator software: a developer framework for driving any desktop app, not a Linux terminal emulator

  Terminator software is an open source desktop automation framework for Windows and macOS. It drives native apps through accessibility APIs, ships as a Rust crate, npm package, Python package, and MCP server, and is published at github.com/mediar-ai/terminator. Not to be confused with the older Linux terminal emulator of the same name.

• Test automation for desktop applications, with a 10ms primary-selector grace period

  Terminator races the primary selector against every alternative in parallel tokio tasks, then gives the primary a 10 millisecond grace window to catch up even when an alternative wins first. The race sits at utils.rs line 2402, the grace at line 2425. That is why a desktop regression suite that lands on Terminator stops silently drifting to weaker selectors between runs.

• Test automation for desktop applications, with a built-in four-file forensics bundle on every step

  Terminator writes four artifacts to disk for every MCP tool call: a JSON request and response log, a regenerated TypeScript replay snippet, a before-screenshot, and an after-screenshot. All land under %LOCALAPPDATA%/mediar/executions/ with 7-day automatic retention. That is a built-in flake-investigation primitive no other desktop automation framework ships. Source: crates/terminator-mcp-agent/src/execution_logger.rs.

• Test automation UI: stop flakes at record time, not run time

  Most test automation UI guides cover run-time flake mitigation: retries, auto-wait, AI self-healing. This page covers the upstream move Terminator

• The automation software for Windows that also runs JavaScript inside your browser tabs

  Most automation software for Windows stops at the OS boundary. Terminator ships a Manifest V3 Chrome extension, Terminator Bridge v0.24.32, that opens a WebSocket to 127.0.0.1:17373 and evaluates JavaScript inside the active tab via chrome.debugger. One script can click a native Windows button and read the DOM of the page behind it.

• The automation software for Windows that can pause and ask the human mid-run

  Most automation software for Windows plays back a recording and fails silently when the UI is ambiguous. Terminator

• The Claude Code MCP server that treats your context window as a budget, not a default

  Every

• The UI automation tool that finds buttons by geometry, not by ID

  Terminator is a UI automation tool with spatial selectors: rightof:, leftof:, above:, below:, near:. Each picks desktop controls by bounding-rectangle geometry against a labeled anchor, so you can target the OK button next to an error message without knowing its AutomationId. Implementation: engine.rs line 1815, NEAR_THRESHOLD = 50.0, MIT licensed.

• The UI test automation tool whose write primitives refuse to lie

  Terminator is a UI test automation tool where every write (type, click, set_value, set_selected) re-reads the element

• The Windows automation program you can watch work: Terminator

  Terminator is the only Windows automation program that draws a click-through, 30% opaque, full-screen status HUD over the live desktop on every click and keystroke. 77/255 alpha, 32-pixel Segoe UI, 300ms minimum display. Source: crates/terminator/src/platforms/windows/action_overlay.rs, 565 lines.

• The Windows automation software that throws out unstable selector text at record time

  Most Windows automation software records whatever the UIA tree returns, then wonders why replays break three steps later. Terminator

• The Windows automation tool whose if: expressions survive a paste from Slack

  Every Windows automation tool has conditional steps. Only Terminator

• Tools for automation testing of desktop application, with a structural diff baked into every action

  Most desktop test tools want you to write assertions per step. Terminator returns a unified diff of the UI tree before and after each action, with element IDs and pixel coordinates already stripped. The stripping logic lives at crates/terminator/src/ui_tree_diff.rs lines 7 and 40. The diff sits on every ClickResult and ActionResult as the uiDiff field.

• TV to use as computer monitor: the coordinate-space gotcha every hardware guide skips

  Hardware blogs cover input lag, OLED burn-in, and chroma 4:4:4. The part nobody writes: the moment your TV becomes display 2, every script, AI agent, and accessibility tool sees a different coordinate space. Here is the exact 95 lines of Rust in Terminator that decide which screen owns each click, and the strict-less-than rule on element.rs:748 that picks the TV over the laptop on a single off-by-one pixel.

• UI Automation for Microsoft Windows: the overlay no other UIA client draws

  Every Microsoft UI Automation client reads the same IUIAutomationElement surface. Terminator is the only one that paints a full-screen click-through status overlay for every action. Window class TerminatorActionOverlay, alpha 77/255, MINIMUM_DISPLAY_MS 300, wrapped in an RAII guard so every click, type, invoke, scroll, and toggle draws it. Source: action_overlay.rs.

• UI automation testing that survives a layout shift: how Terminator diffs the accessibility tree without coordinates

  Most UI automation testing breaks the moment a button moves 100 pixels. Terminator diffs the OS accessibility tree with the bounds rectangles stripped out, so a Reply button that drifts 118px down between runs produces zero diff and your test still passes. Source: ui_tree_diff.rs.

• UI automation tests that locate elements by geometry: the five spatial selectors inside Terminator

  Most UI automation tests lock onto IDs, names, or XPaths that shift the moment the layout changes. Terminator ships five spatial selectors (rightof:, leftof:, above:, below:, near:) that find an element by its position relative to a visible anchor. Near uses a hard-coded 50.0 pixel Euclidean radius.

• UI automation tools in 2026: every listicle still ships a GUI; Terminator ships 35 MCP tools

  The top eight ranking pages for ui automation tools list Selenium, Playwright, TestComplete, Ranorex, Katalon, UiPath. All of them are humans driving a GUI. Terminator inverts that: 35 MCP tools exposed from one Rust binary, so Claude Code, Cursor, VS Code, and Windsurf can click, type, and navigate desktop apps directly. One install line. Source at crates/terminator-mcp-agent/src/server.rs.

• UI automation, Selenium-shaped, recorded as semantic events instead of keystrokes: Terminator

  Selenium IDE records clicks and keystrokes inside one browser tab. Terminator

• UI test automation tools: the 25-variant selector enum every listicle is missing

  Every 2026 roundup of ui test automation tools lists browser-only frameworks or closed-source image-based suites. None include a free, code-first tool that targets any native desktop app via the OS accessibility tree. Terminator

• UiPath desktop automation for developers: a selector grammar with operator precedence, not an XML editor

  Most explainers about UiPath desktop automation describe XML selectors with wnd and ctrl tags configured inside Studio

• What is an MCP server? A real one, opened up in the editor, not the abstract definition

  Every top result for this query defines MCP the same way: a protocol that lets AI call tools. Nobody opens an actual MCP server and shows the dispatch function. This page does. We look at Terminator

• What is UI automation? It is pattern invocation, not pixel clicking

  UI automation is a program that drives an application through its accessibility tree. In Terminator, invoke() is 22 lines of Rust that dispatch a UIA InvokePattern through COM, no mouse involved. click() is a 5-phase cascade that falls back to a real SendInput mouse event. Most guides describe the outcome; this one reads the source.

• Why accessibility APIs beat OCR and pixel matching for OS-level automation

  OCR and pixel matching identify a button by what it looks like to the camera. Accessibility APIs identify it by what the developer named it in code. The three places this matters most are latency, stability, and i18n. Terminator

• Windows automation in Python with a real visual debugger built in

  Every other Python library for Windows automation runs your script blind. Terminator

• Windows automation scripting, inverted: the AI drives, the runtime writes the script

  Every other guide hands you a PowerShell or AutoHotkey template. Terminator flips it. An AI coding agent drives the desktop through an MCP server, and execution_logger.rs transcribes every tool call into a paired JSON log and replayable TypeScript snippet with before/after screenshots, saved under %LOCALAPPDATA%\\mediar\\executions\\ for 7 days.

• Windows automation with Python that runs PowerShell inline, no subprocess needed

  Terminator

• Windows program automation: the three launch backends behind one open_application call

  Power Automate Desktop, AutoHotkey, and PyAutoGUI all assume the program is already open. Terminator

• Windows script automation: one MCP tool, six scripting engines, one desktop SDK

  PowerShell, Node, Bun, TypeScript, Python, cmd, bash. One MCP tool (run_command) drives all of them and injects a desktop SDK as a global so a single Windows script can shell out to the registry and click a WPF button in the same execution. Source: terminator-mcp-agent/src/utils.rs:963.

• Windows software automation as a query language: selectors with &&, ||, !, >>, and spatial anchors

  Most Windows software automation tools record clicks or pin to X,Y coordinates. Terminator turns the desktop into a query surface: role:Button && name:Save, window:Login >> role:Edit && rightof:(name:Username), !name:Cancel. This is a tour of the selector grammar, the Shunting Yard parser that compiles it, and the 50.0-pixel geometric filter that powers rightof: and near:. Source: crates/terminator/src/selector.rs and platforms/windows/engine.rs.

• Windows task automation: the deferred click capture that makes recording a real human workflow survive the modal it just opened

  Recording a Windows workflow is a race against the UI the click is about to change. Terminator