# Control protocol Start the app with `--control ` (or `DESIGNCRAFT_CONTROL_PORT`). The server listens on `127.0.0.1` only and speaks JSON lines: one request object per line, one reply per line. ```json {"id": 1, "method": "engine.execute", "params": {"command": "frame.create", "params": {"rect": [36, 36, 300, 200], "content": "text"}}} {"id": 1, "ok": true, "result": {"id": 10, "story": 11}} ``` | Method | Params | What it does | |---|---|---| | `engine.execute` / `ui.menu.invoke` | `{command, params}` | Run any engine or UI command (see `engine.commands`). `ui.menu.invoke` without `params` acts like choosing the menu item: a command labelled "…" that takes parameters opens its dialog | | `engine.commands` | — | Every command: id, label, menu path, shortcut, params doc, enabled | | `document.inspect` | — | Pages, spreads, items, stories (overset), styles, swatches, selection | | `ui.inspect` | — | Tool, UI state, view (zoom/origin), canvas rect, perf | | `ui.menu.list` / `ui.tool.list` | — | Menu tree / Tools panel groups | | `ui.tool.select` | `{tool}` | Select a tool (`selection`, `type`, `rectangleFrame`, …) | | `ui.pointer` | `{events:[{kind: down\|drag\|up\|move\|doubleclick, x, y, space?: "screen"\|"canvas"}], mods?}` | Drive the active tool through the same code path as the mouse | | `ui.key` / `ui.text` | `{key, shift?, alt?, cmd?}` / `{text}` | Synthetic keyboard input (typing into a text frame) | | `ui.move` / `ui.click` / `ui.drag` | screen points | Real egui pointer input — reaches every widget, menu and panel | | `ui.set` | `{brightness?, panel?, rulers?, guides?, frameEdges?, baselineGrid?, textThreads?, screenMode?, zoom?, page?, fit?}` | UI state | | `ui.dialog.open` | `{id, fields?}` | Open a dialog by id (e.g. `paragraphStyleOptions` with `{name, section}`) | | `ui.dialog.set` / `ui.dialog.confirm` / `ui.dialog.cancel` | `{field, value}` | Fill and confirm the open dialog | | `ui.resize`, `ui.focus` | | Window control | | `ui.screenshot` | `{path?}` | PNG of the whole window | | `ui.render` | `{path?, page?, scale?, bleed?}` | Render a page headlessly (PNG; base64 if no path) | | `app.open` / `app.save` / `app.export` / `app.quit` | | Files | Headless window screenshots (locked screen, hidden window): `cargo run -p designcraft-ui-egui --example ui_shot -- script.jsonl`, where each line is one of the requests above, `{"shot": "/abs/out.png"}` or `{"steps": n}` (renders the whole UI offscreen with wgpu). The MCP server (`designcraft-cli mcp`) wraps the same methods for Claude and other agents.