wtf is open-controller?

open-controller is a plugin for OpenCode, a coding assistant tool, that lets an AI agent directly control a Windows computer instead of just editing code. It works by connecting to a separate helper program called windows-mcp, which does the actual work of clicking, typing, and reading the screen. Once installed, the AI agent gains the ability to click, type, scroll, drag, and switch between application windows on the desktop, take and read screenshots, read and write files and folders, run PowerShell commands, read and change Windows Registry entries, list or stop running programs, get or set clipboard content, send Windows notifications, fetch and read web pages, and select or edit multiple on screen elements at once. In practice this means you could ask the AI agent to do things like open a program, fill in a form, take a screenshot to check the result, or run a script, all through plain instructions rather than code. Setting it up requires Node.js, OpenCode itself, and Python, since the underlying windows-mcp helper is a Python package installed separately. Installation is a matter of adding the plugin through npm and enabling it in OpenCode's configuration file, then installing the Python package with pip. A notable design detail is that the plugin ships a written set of instructions that it automatically inserts into the AI agent's system prompt every session, covering common mistakes such as the agent acting on the wrong window, typing long text too quickly, or misreading a timeout message as a failure when it was actually just a slow operation. The plugin also raises the normal 30 second timeout most tools use up to 20 hours by default, since some UI automation actions legitimately take a long time, though this can be changed with an environment variable. Starting from a later version, the plugin checks for and installs its own updates automatically each time OpenCode starts, unless this is turned off.