Every listicle for UI automation tools still ships a GUI. Terminator ships 35 MCP tools.

Traditionally, any software that drives a graphical interface without a human moving the mouse. The commodity list has not moved in a decade: Selenium, Playwright, Cypress, Appium, TestComplete, Ranorex, Katalon, UiPath, Power Automate Desktop, Automation Anywhere, LambdaTest, BrowserStack, Applitools. Every one of them either ships an IDE for a human tester or a recorder that writes a script. The category quietly grew a second branch around 2024 when the Model Context Protocol landed: UI automation tools that are not apps at all, just a bundle of callable tools the AI coding assistant holds in its context window. Terminator is in that second branch. It exposes 35 tools over MCP, and the primary user is Claude Code or Cursor, not a QA engineer with a mouse.

35. Run grep -c '#\[tool(' crates/terminator-mcp-agent/src/server.rs against the repo and you will count exactly that many annotations. The list includes get_window_tree (scan the accessibility tree of a running process), click_element, type_into_element, press_key, open_application, navigate_browser, scroll_element, select_option, set_toggled, capture_screenshot, validate_element, wait_for_element, execute_sequence (chain multiple tool calls into one deterministic block), zoom, mouse_drag, read_file, write_file, edit_file, grep_files, glob_files, copy_content, stop_all_workflows, typecheck_workflow, and the agentic loop run_ai_agent. That last one is the Gemini 2.5 Computer Use model wired in as a fallback when deterministic selectors are not enough. No other tool on the top-ranking lists does this.

Because the primary new user of a UI automation tool is no longer a QA engineer, it is a coding agent. Claude Code, Cursor, VS Code's Agent mode, and Windsurf all speak MCP. When you install an MCP server, the agent sees the tool list in its context. The agent picks which tool to call based on the user's natural-language goal: find the OK button, type my password, check the error dialog, close the modal. The agent does not need a GUI. It needs a list of tools with clear names, clear arguments, and return values it can read back. Terminator is designed around that interface. Playwright MCP exists for browser-only control. Terminator MCP runs against the whole desktop.

claude mcp add terminator "npx -y terminator-mcp-agent@latest". Restart Claude Code. Type a goal: open Chrome, navigate to github.com, find the issues tab, filter by label. Claude will plan the tool calls, resolve selectors against the live accessibility tree, and execute. No test script, no fixture, no page object. The same binary works for Cursor and VS Code through their MCP config files, and for any custom client via the stdio transport.

Playwright has an MCP server, but it is browser-only and runs headless by default. Puppeteer has community MCP wrappers, same limitation. Microsoft's Playwright Test recorder is a standalone app, not an MCP tool. UiPath has API access but no official MCP bridge. Power Automate Desktop has no MCP server. That leaves Terminator as the only production-ready MCP UI automation tool that covers browsers AND native Windows apps in the same install. The combined coverage is what makes it useful for agent workflows that span more than a single tab.

Both, and the line is blurring. For QA, the 35 MCP tools include validate_element, wait_for_element, and typecheck_workflow which read cleanly as an assertion library. For production automation, execute_sequence lets you pre-record a deterministic path and only fall back to the AI agent loop when the UI shifts. The workflow recorder in crates/terminator-workflow-recorder captures a human doing a task and produces a replayable sequence. The MCP server then runs that sequence with the AI as the recovery layer.

Terminator's Rust core compiles on macOS via the AX API and on Linux via AT-SPI2. The published npm, pip, and MCP binaries are Windows-only as of April 2026. The llms.txt in the repo is explicit about it. If you need desktop automation on macOS today, the core Rust crate works but the MCP agent's tools list is Windows-first. The roadmap tracks macOS MCP parity.

MIT. The repo is at github.com/mediar-ai/terminator. Fork it, ship it, embed it in your own product, commercial or not. Most of the commercial UI automation tools on the SERP lists are proprietary with per-seat pricing or per-bot pricing that starts above $1000/year. The open-source ones on the list (Selenium, Appium, Playwright) are browser-first. Terminator is the only MIT-licensed UI automation tool in 2026 that covers the full Windows desktop AND exposes itself as an MCP server.