cursor-cdp-control

# Cursor CDP Control Control Cursor IDE via `opencli cursor` commands using Chrome DevTools Protocol (CDP). All commands auto-detect window type: - **Agent window** - Cursor 3 standalone Agent window (title = "Cursor Agents") - **Editor window** - Traditional editor with embedded Chat panel ## Prerequisites ### 1. Launch Cursor with CDP port enabled Close all Cursor instances first, then relaunch: ```bash # Windows "C:\Program Files\cursor\Cursor.exe" --remote-debugging-port=9226 # macOS /Applications/Cursor.app/Contents/MacOS/Cursor --remote-debugging-port=9226 ``` **Notes:** - Must close all Cursor windows before launching (Electron single-instance lock) - Default port is `9226`; customize with `--port` argument - Do not use the `cursor` CLI tool (it invokes cli.js, not the GUI) ### 2. Install plugin dependencies ```bash cd ~/.opencli/plugins/ksh-plugin-cursor && npm install ``` ### 3. Verify connection ```bash opencli cursor status ``` ## Command Reference | Command | Description | Example | |---|---|---| | `cursor status` | CDP connection status | `opencli cursor status` | | `cursor list` | List all windows (Agent/Editor) | `opencli cursor list` | | `cursor new-chat` | Open a new Agent/Chat tab | `opencli cursor new-chat` | | `cursor project` | List available projects (Agent only) | `opencli cursor project` | | `cursor project-switch` | Switch project by fuzzy match (Agent only) | `opencli cursor project-switch "llm-wiki"` | | `cursor send` | Send prompt and wait for reply | `opencli cursor send "refactor this code"` | | `cursor read` | Read current Chat content | `opencli cursor read --limit 10` | | `cursor history` | Browse chat history sessions | `opencli cursor history "test"` | | `cursor model` | List available models | `opencli cursor model` | | `cursor model-switch` | Switch model by fuzzy match | `opencli cursor model-switch "sonnet"` | | `cursor screenshot` | Capture screenshot | `opencli cursor screenshot chat` | See [references/commands.md](references/commands.md) for detailed parameters. ## Window Type Differences All commands auto-detect, no manual specification needed: | Component | Agent Window | Editor Window | |---|---|---| | Input box | TipTap ProseMirror | aislash-editor | | History | Sidebar, grouped by project, always visible | Click button to show popup panel | | Chat area | `.agent-panel` | composer container | | Model picker | Shared `.ui-model-picker__trigger` | Shared | | Project picker | Supported (`.ui-select-trigger`) | Not supported | Use `--window N` to target a specific window (N matches `cursor list` idx, 1-based). Default 0 auto-selects Agent window. ## Typical Usage ```bash # List all windows opencli cursor list # Open a new Agent tab opencli cursor new-chat # Send to a specific project (auto-finds window by project name, works in both Agent & Editor) opencli cursor send "analyze this code" --project "mcd" # Switch project in Agent window (Agent only) opencli cursor project-switch "llm-wiki" # Send a prompt (auto-selects Agent window) opencli cursor send "analyze this code" # Browse Editor window history opencli cursor history "refactor" --window 3 # Switch model and take screenshot opencli cursor model-switch "sonnet" opencli cursor screenshot chat ``` ## Technical Details Uses CDP HTTP API to discover window targets, then communicates via WebSocket: - `Runtime.evaluate` - DOM queries and data extraction - `Input.insertText` / `Input.dispatchMouseEvent` / `Input.dispatchKeyEvent` - Simulated user interaction - `Page.captureScreenshot` - Headless screenshot to file Auto-detects window type by target title, applies type-specific DOM selectors. Non-invasive, no native hooks.