Baguette

Bon appétit.

Headless iOS Simulator manager + host-side input injection for iOS 26.

A single Swift CLI — baguette — plus a self-contained web UI that gives you full headless control of an iOS simulator without opening Xcode or Simulator.app. Boot devices, stream their screens at 60 fps, dispatch taps / swipes / multi-finger gestures / system gestures / keyboard / hardware buttons, tail the unified log, inspect the accessibility tree, take screenshots and recordings, and — as of 0.1.72 — pipe your Mac webcam into the simulator's camera APIs.

Demo

https://github.com/user-attachments/assets/e904413f-16bb-4b3d-86d5-162333403cee

https://github.com/user-attachments/assets/c49c9f4b-0e4b-47ea-9272-3223b1ac7739

https://github.com/user-attachments/assets/65dc62ee-f0c7-48fb-9c57-5bd267c8c02f

The raw clip lives at assets/demo.mp4 — drag it into a GitHub web edit of this README to upload as a CDN-hosted video and replace the line above with the auto-generated URL.

• Frame streaming — MJPEG or H.264 / AVCC over stdout or WebSocket. Runtime-tunable bitrate / fps / scale. In-browser recording (MP4) composites the bezel + screen + gesture overlays into one file.
• Host-HID input — taps / swipes / streaming 1- and 2-finger gestures / pinch / pan / scroll / Mac keyboard / hardware buttons (home, lock, power, volume, action, plus Apple Watch's digital crown + side button) — all through SimulatorKit's private symbols with the iOS-26 calling conventions. The iOS-26 streaming-touch + edge-gesture path uses IOHIDDigitizerDispatch so home-indicator swipes, app-switcher drags, and Notification Center / Lock Screen pull-downs all fire the real iOS recognizers live. No dylib injection on this path; no DYLD_INSERT_LIBRARIES to manage.
• Camera (new in 0.1.72) — pipe a Mac webcam directly into the iOS simulator's AVCaptureVideoPreviewLayer, AVCapturePhotoOutput, and UIImagePickerController. Pick a camera in the browser's Camera card, click Start, the iOS app sees real frames. One ObjC dylib (VirtualCamera.dylib, vendored from asc-pro/SimCam) loaded into every sim-launched app via DYLD_INSERT_LIBRARIES; baguette pumps BGRA frames through a shared-memory ring buffer. See docs/features/camera.md.
• Device orientation — baguette orientation --udid <X> portrait rotates a booted simulator. Wire JSON + a one-click rotate button on the focus-mode toolbar. Fires a GSEventTypeDeviceOrientationChanged mach message at PurpleWorkspacePort, bypassing SimulatorKit's NSView path so the host stays headless.
• Accessibility tree — baguette describe-ui returns the on-screen AX tree as JSON (per-node role, label, value, identifier, frame in device points). Hit-test mode (--x --y) returns the topmost node under a coordinate. Powered by the private AccessibilityPlatformTranslation framework with a bridgeTokenDelegate we install ourselves.
• Live unified-log stream — baguette logs --udid <X> streams os_log output to stdout; WS /simulators/:udid/logs does the same to the browser's Logs panel. Predicate / bundle-id filters.
• Standalone web UI — baguette serve opens http://localhost:8421/simulators with a list page, a focus-mode per-device view, a sidebar stream view, the Camera card, an Accessibility inspector overlay, a Logs panel, and in-browser recording. All wrapped by a small JS SDK (Resources/Web/baguette/) — const sim = await Baguette.use({…}); sim.mount(container); — that hangs each part (screen, buttons, keyboard, …) off one Simulator instance.
• Device farm — http://localhost:8421/farm renders every booted simulator in a wall / grid / list with filtering + sorting. Click a tile to focus it for full-quality streaming + input through the same pipeline as the CLI.
• TDD non-negotiable, layered, mock-injected — bounded-context Domain / Infrastructure / App split; 460+ Swift Testing cases backed by auto-generated MockXxx fakes for every external port (Input, Screen, Accessibility, LogStream, Chromes, DeviceHost, Subprocess, CameraCapture, VideoCapture, CameraFrameSink, SimulatorInjection, Cameras). swift test requires no simulator at all.

Install

brew install baguette

Apple Silicon only. Requires Xcode 26 — baguette links against private SimulatorKit / CoreSimulator frameworks shipped with Xcode.

Quickstart

# Start the web UI
baguette serve

# Single-device dashboard — list, boot/shutdown, per-device stream pages
open http://localhost:8421/simulators

# Device farm — every booted simulator side-by-side, click to focus
open http://localhost:8421/farm

/simulators lists every simulator on the machine with Boot / Shutdown buttons; click any booted device to open its focus-mode page — full-window live stream, DeviceKit-sourced bezel, top toolbar with Camera / Accessibility / Logs / Home / Screenshot / App-switcher controls, and a sidebar-view jump button.

/farm is the multi-device control surface. See Device farm below.

Headless from the terminal works too:

baguette list
baguette boot --udid <UDID>
baguette tap --udid <UDID> --x 219 --y 478 --width 438 --height 954

Build from source

make           # release build via ./build.sh
swift test     # run the test suite

Hybrid build: SPM fetches dependencies (ArgumentParser, Mockable, Hummingbird, HummingbirdWebSocket); swiftc compiles everything with an Objective-C bridging header targeting arm64e-apple-macos26.0, linking CoreSimulator, SimulatorKit, IOSurface, VideoToolbox, CoreGraphics, ImageIO from Xcode's private frameworks.

CLI

baguette <command> [options]

  # Lifecycle
  list [--json]                              List devices (default + custom sets;
                                             --json emits {"running":[…],"available":[…]})
  boot     --udid <UDID>                     Boot headlessly
  shutdown --udid <UDID>                     Shutdown
  orientation --udid <UDID>                  Rotate the booted simulator
              <portrait|landscape-left|       (GSEvent over PurpleWorkspacePort —
               landscape-right|portrait-      no NSView, host stays headless)
               upside-down>

  # Frames + screenshots
  stream     --udid <UDID> [--fps 60] [--format mjpeg|avcc]
                                             Stream frames on stdout
  screenshot --udid <UDID> [--output <path>] [--quality 0.85] [--scale 1]
                                             One-shot JPEG (defaults to stdout)

  # Accessibility + logs
  describe-ui --udid <UDID> [--x <px> --y <px>] [--output <path>]
                                             Dump on-screen accessibility tree as
                                             JSON; frames in DEVICE POINTS so
                                             they pipe straight back into a tap.
  logs --udid <UDID> [--level info|debug|default]
                     [--style default|compact|json|ndjson|syslog]
                     [--predicate <NSPredicate>] [--bundle-id <id>]
                                             Stream os_log output. Levels are
                                             the three the iOS-runtime accepts.

  # Long-lived gesture pipe
  input --udid <UDID>                        Read newline-delimited JSON
                                             gestures from stdin

  # Web UI — single-device dashboard + multi-device farm + Camera card +
  # Accessibility inspector + Logs panel + in-browser recording.
  serve [--port 8421] [--host 127.0.0.1] [--device-set <path>]

  # DeviceKit chrome / bezel data
  chrome layout    --udid <UDID> | --device-name "iPhone 17 Pro"
  chrome composite --udid <UDID> | --device-name "iPhone 17 Pro"

  # One-shot gestures — same HID path as `input`, one gesture per
  # invocation. Coordinates are in DEVICE POINTS; `width` / `height`
  # are the simulator's screen size in points.
  tap        --udid … --x … --y … --width … --height … [--duration 0.05]
  double-tap --udid … --x … --y … --width … --height …
                                             [--interval 0.05] [--duration 0.08]
  swipe      --udid … --startX … --startY … --endX … --endY …
                                             --width … --height …
  pinch      --udid … --cx … --cy … --startSpread … --endSpread …
                                             --width … --height …
  pan        --udid … --x1 … --y1 … --x2 … --y2 … --dx … --dy …
                                             --width … --height …

  # Keyboard (single keystroke or typed string)
  key  --udid … --code <KeyA..Z|Digit0..9|Enter|Escape|Backspace|Tab|Space|
                       Arrow*|punctuation>
                       [--modifiers shift,control,option,command] [--duration 0.2]
  type --udid … --text "<US-ASCII string>"

  # Hardware + virtual buttons. Phone: home / lock / power / volume-up /
  # volume-down / action / app-switcher / swipe-to-app-switcher /
  # swipe-to-home / pull-down-to-lock-screen / pull-down-to-notification-center.
  # Watch: digital-crown / side-button / left-side-button.
  press --udid … --button <name> [--duration <sec>]

baguette serve — the web UI

baguette serve --port 8421
# [baguette] listening on http://127.0.0.1:8421/simulators

Open http://localhost:8421/simulators in any browser. You get the device list (RUNNING / AVAILABLE), Boot / Shutdown buttons, and a per-device focus-mode page at /simulators/<UDID> with live frames, gesture input, the DeviceKit-sourced bezel, a top toolbar (Camera / Accessibility / Logs / Home / Screenshot / App-switcher / Rotate), floating Camera + Accessibility control cards, and an in-browser MP4 recorder. A sidebar-view variant is reachable from the bottom-left toggle on the focus page.

The HTML is editable on disk — Sources/Baguette/Resources/Web/sim.html opens directly in any browser via file:// (preview mode), and points to its sibling .js files. Set BAGUETTE_WEB_DIR to override the served root for live-iteration without rebuilding.

Routes (single resource tree, no /api/ prefix)

One bidirectional WebSocket per stream

The same WS carries everything for a viewing session:

• Server → Browser — encoded binary frames (one per WS message).
  • MJPEG: raw JPEG bytes per frame.
  • AVCC: 1-byte tag + payload — 0x01 avcC description, 0x02 keyframe, 0x03 delta, 0x04 JPEG seed (renders before H.264 IDR lands).
• Browser → Server — text JSON, one line per message:
  • Stream control: {"type":"set_bitrate","bps":N} / {"type":"set_fps","fps":N} / {"type":"set_scale","scale":N} / {"type":"force_idr"} / {"type":"snapshot"}.
  • Gesture input: same wire format as baguette input (see below).

No /event POST, no UDID-keyed registry — the WS handler closure owns the live stream + simulator handle for the duration.

Device farm

baguette serve
open http://localhost:8421/farm

A multi-device dashboard for the booted simulators on the host. Every device renders in a single page; the same WebSocket pipeline that powers /simulators/:udid drives every tile.

What it does

• Three view modes — Grid (compact thumbnails), Wall (large tiles with bezels), and List (one-row-per-device with metadata). Toggle from the header.
• Filter and sort — by device family, OS version, run state. The rail on the left holds filter state across view changes.
• Click to focus — clicking any tile re-parents its <canvas> into a full-quality focused pane on the right. The thumbnail keeps streaming at low bitrate; only the focused tile pays for full-rate frames. No separate m