Automation for Windows your AI coding assistant can actually drive

Power Automate Desktop is a low-code canvas for human operators to click together. AutoHotkey is a hotkey and scripting language. Terminator is a developer framework, closer to Playwright for the whole OS. You write code (Rust, TypeScript, or Python) or you let an AI coding assistant call its MCP tools directly. Selectors are UIA-based strings like role:Button && name:Save, scoped by process. No drag-and-drop canvas, no vendor lock-in, MIT licensed.

Exactly 35, registered via #[tool(...)] attributes inside crates/terminator-mcp-agent/src/server.rs. They include click_element, type_into_element, press_key, press_key_global, invoke_element, activate_element, mouse_drag, scroll_element, select_option, set_selected, set_value, capture_screenshot, get_window_tree, get_applications_and_windows_list, navigate_browser, open_application, validate_element, wait_for_element, highlight_element, stop_highlighting, hide_inspect_overlay, delay, ask_user, gemini_computer_use, execute_sequence, read_file, write_file, edit_file, copy_content, glob_files, grep_files, typecheck_workflow, stop_execution. Grep the source: grep -c '#\[tool(' crates/terminator-mcp-agent/src/server.rs returns 35.

When an AI agent clicks a button, it needs to know what changed. The naive approach is to re-fetch the full window tree after every action, which is slow and floods the agent context with unchanged nodes. Terminator snapshots the tree before the action, runs the action, snapshots the tree after, and returns only the diff. The tool description strings literally tell the model: 'Use ui_diff_before_after:true to see changes (no need to call get_window_tree after).' See server.rs line 2152 for the type_into_element tool's description.

UIA trees contain volatile identifiers like runtime ids and bounding rectangles that shift on every repaint, so a naive diff would surface dozens of cosmetic changes for every click. Terminator strips those fields in crates/terminator/src/ui_tree_diff.rs. preprocess_tree removes id and element_id from the JSON form. remove_ids_and_bounds_from_compact_yaml strips '#abc123' identifiers and 'bounds: [x,y,w,h]' tuples from the compact YAML form. After preprocessing, a text diff via the similar crate shows only semantic changes.

Speed and determinism. Per Terminator's llms.txt, UIA-based automation runs at CPU speed rather than LLM inference speed, which is roughly 100x faster, and it succeeds over 95 percent of the time on deterministic tasks. Screenshot-plus-vision agents have to re-infer the same layout on every step, drift when icons change, and cost money per action. Terminator reserves AI for error recovery, not every click.

One command. Run claude mcp add terminator 'npx -y terminator-mcp-agent@latest' in a terminal and Claude Code registers all 35 tools. The agent has two identical one-liners in the MCP config for Cursor, VS Code, and Windsurf, documented in the mediar-ai/terminator README. After that, any prompt like 'open Excel, paste this formula in cell A1, save the file' resolves through the MCP tools instead of screenshots.

Both. The Windows backend talks to the UI Automation COM API (IUIAutomation, introduced in Windows 7 and stable across every shipping version since). The MSI installer and prebuilt npm binary target Windows 10 and 11, both x64 and ARM64. The Rust core also has macOS and Linux backends (AX API and AT-SPI2), though the npm, pip, and MCP packages currently ship Windows binaries only.

Yes. The ui_diff_before_after response includes the set of added, removed, and modified nodes between the two tree snapshots, rendered in the compact YAML form after ID and bounds stripping. If the click opened a dialog, the diff contains the new [Window] and its children. If the click was eaten by a disabled control, the diff is empty. The agent decides what to do next from that text alone.