kwcapture — fast screen capture on Wayland (KDE Plasma)

PIL.ImageGrab runs at ~2 fps on Wayland (it shells out to spectacle) and mss captures XWayland, which is black for native Wayland windows. kwcapture talks to the KWin compositor directly: ~40 fps at 2560×1440, real pixels, ~26 ms from "give me a frame" to a JPEG ready for a vision model. Whole screen, a region, one window by name or id, or one monitor by name or id — list_windows() and list_monitors() give you both.

Scope: KDE Plasma only. The kw in the name is the design. kwcapture speaks org.kde.KWin.ScreenShot2 and nothing else; compositors that only offer wlr-screencopy / ext-image-copy-capture-v1 are out of scope, and it will say so rather than half-work on them.

method                        fps      median latency
mss (XWayland)              1113.8      0.9 ms   <-- BLACK, useless on Wayland
PIL.ImageGrab (spectacle)      2.7    357.1 ms
kwcapture raw BGRA view       38.4     25.8 ms   real pixels, zero copy
kwcapture RGB full res        37.9     26.0 ms
kwcapture JPEG 1280 wide      35.6     28.0 ms   includes downscale + encode
kwcapture one 692x440 window 185.4      4.7 ms   per-window capture, not clipped

import kwcapture as K

cap  = K.Capture()                      # starts the helper, first frame in ~50 ms
arr  = cap.grab()                       # (H, W, 4) BGRA view into shared memory
rgb  = cap.shot(width=1280)             # (720, 1280, 3) uint8, contiguous
blob = cap.shot_jpeg(1280, quality=85)  # bytes for a vision API
cap.stats()                             # {'grab_ms': 20.1, 'total_ms': 27.8, ...}
cap.close()

Per-window capture — the window handle is KWin's own id, so no geometry juggling and no cropping by whatever is on top of it:

for w in K.list_windows():                # name + handle of every capturable window
    print(w.id, w.name, w.app_id, w.geometry, w.active)   # active == has focus

win = K.Capture(window="Kate")            # by name, caption or app id …
kon = K.Capture(window="{e0b1aab4-4e10-…}")   # … or by the handle it came with
act = K.Capture(active_window=True)       # whatever has focus
win.grab()                                # 692x440 KCalc window: ~5 ms, ~185 fps
win.shot_jpeg(quality=90)                 # just that window

K.active_window()                         # the focused window, as a Window
K.active_window_id()                      # … or just its KWin handle (~8 ms)

A minimised window returns stale pixels — KWin stops rendering it, so every grab repeats the same buffer and nothing tells you. Ask explicitly, or opt in to a per-grab check:

win = K.Capture(window="Kate", stale_check=True)   # opt-in: one throttled query per grab
frame = win.grab()
if win.stale_frame:             # the window is minimised -> frame is an old snapshot
    ...
win.window_minimized            # ask KWin right now (a D-Bus round trip; not per frame)

stale_check is off by default because the check costs a round trip; without it stale_frame stays False, which means unknown rather than fresh.

Install

pip install kwcapture                     # PyPI
pip install "kwcapture[fast]"             # + OpenCV: ~8x faster resize/encode

# straight from GitHub (builds the helper for your machine):
pip install "kwcapture @ git+https://github.com/tjandrasg/kwcapture.git"

# or the prebuilt wheel straight from the release page:
pip install https://github.com/tjandrasg/kwcapture/releases/download/v0.5.0/kwcapture-0.5.0-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl

Needs KDE Plasma with KWin on Wayland. The PyPI wheel ships the small native helper prebuilt for x86-64 Linux (glibc ≥ 2.17), so no compiler is required; if your distro or architecture has no wheel, pip compiles it on install, which needs a C compiler and libsystemd/wayland-client headers (on Debian/Ubuntu: sudo apt install build-essential libsystemd-dev libwayland-dev). Only numpy is a runtime dependency; the helper links libsystemd and libwayland-client, which any KDE desktop already has.

Then, in the graphical session:

kwcapture doctor          # checks session, helper, KWin authorisation, capture

The first Capture() writes the one file KWin needs to authorise us (~/.local/share/applications/io.github.kwcapture-<hash>.desktop) — no manual setup step. Extras: pip install "kwcapture[fast]" (OpenCV: ~8x faster resize/encode) or "kwcapture[pil]" (Pillow).

Why a native helper (the interesting bit)

KWin exports org.kde.KWin.ScreenShot2 over D-Bus: you hand it a file descriptor and it writes raw ARGB32-premultiplied (BGRA in memory) pixels into it. But it authorises callers by /proc/<pid>/exe: that binary must match the Exec= of a .desktop file declaring X-KDE-DBUS-Restricted-Interfaces=org.kde.KWin.ScreenShot2 — the trick Spectacle uses. A Python interpreter can never satisfy that, so:

kwcapture.py  ──poke──▶  helper (serve)  ──D-Bus──▶  kwin_wayland
      ◀──mmap ring────        ◀──pipe(fd)────────

The helper keeps one D-Bus connection open and publishes frames into a shared-memory ring, so a grab is a byte to a FIFO, a spin on a sequence counter, and a zero-copy numpy view: no process per frame, no serialisation, no PNG round trip. KWin sends the reply before the pixels land, so EOF on the pipe — not the reply — is the frame boundary.

CLI

kwcapture doctor [--fix] [--build]     # is everything OK? (--fix also authorises)
kwcapture setup                        # build the helper + authorise it
kwcapture install-desktop [--uninstall]# manage the KWin authorisation file
kwcapture screens                      # DP-1 2560x1440 @164.69Hz pos 0,0 scale 1
kwcapture windows                      # every capturable window: handle, name, app, size
kwcapture windows -f kcalc --json      # filter by name/app id; machine-readable
kwcapture grab -o shot.png --width 1280
kwcapture grab -o frame.bgra --raw     # pure BGRA bytes
kwcapture grab --window Kate -o kate.png
kwcapture grab --window '{e0b1aab4-4e10-…}' --decoration -o win.png
kwcapture grab --active-window -o focused.png   # the window that has focus
kwcapture demo --frames 60 --save f.png
kwcapture bench --frames 200

kwcapture windows looks like this:

WINDOW HANDLE (id)                      NAME                          APP_ID              SIZE    POSITION  FLAGS
{e0b1aab4-4e10-47fd-ae46-77a4ee6bc4d8}  build-pgo : llama-server      org.kde.konsole  2560x1394      +0,+0  maximized
{2c14f294-93ea-48bf-bbea-f501d59f6fe5}  KCalc                         org.kde.kcalc      692x468  +1563,+738

Python API

Capture(...) arguments:

argument	meaning
screen="DP-1"	which output (default: active screen)
area=(x, y, w, h)	capture a region instead of a whole output
workspace=True	capture the entire virtual desktop
window="Kate"	capture one window: a handle from list_windows(), or its caption / app id
active_window=True	capture the window that has focus
cursor=True	include the hardware cursor
decoration=True	include window decorations (and their shadow); off = client area only
slots=4	ring depth — how many frames until a returned view is overwritten
depth=2	requests in flight to KWin (2 is optimal; more adds latency, not fps)
fps=30	continuous capture mode; read with latest() for zero-latency access
idle_exit=300	helper quits when untouched for this many seconds
binary=…, allow_build=…	override / disable locating-or-building the helper
install_desktop=False	don't touch ~/.local/share/applications
verbose=True	report build/authorisation steps

Methods: grab(rgb=False, copy=False, fresh=True), latest(), shot(width=…, resample=…), shot_png(), shot_jpeg(), stats(), geometry(), screen_name, target, window_id/window_name, bench(frames), restart(), close() (also a context manager). Module helpers: list_screens(), list_windows(), find_window(), is_window_handle(), to_rgb(), resize(), png_bytes(), jpeg_bytes(), default_shm_path(), unique_shm_path(), find_binary(), install_desktop_file(). grab() / shot() also take the target keys (window=, screen=, …) and keep one helper process alive per distinct target.

Notes:

• grab() returns a read-only view of the shared ring, overwritten after slots further frames — use copy=True (or .copy()) to keep a frame, or shot()/to_rgb(). The ring is mapped read-only, so a view can never write into it, and a view stays readable after close() — frozen at the frame that was current when you closed.
• latest() returns the newest published frame without asking KWin for one.
• hide_caller_windows=True by default: KWin hides the capturing process' own windows, so your overlay/terminal does not end up in the shot (--no-hide-caller to disable).
• Each Capture gets its own ring file in $XDG_RUNTIME_DIR (see unique_shm_path()), so any number of them can run at once — one per monitor, one per window. The helper also dies with its client (PR_SET_PDEATHSIG).

Per-window capture

import kwcapture as K

for w in K.list_windows():                 # every window KWin considers capturable
    print(f"{w.id}  {w.name}  {w.app_id}  {w.width}x{w.height}+{w.x}+{w.y}")

win = K.Capture(window="Kate")             # caption, app id, desktop file or handle
try:
    frame = win.grab()                     # (H, W, 4) BGRA, same as a screen grab
except K.WindowGone:                       # someone closed the window
    win.close()
    win = K.Capture(window="Kate")         # a new window with the same name

focused = K.Capture(active_window=True)    # whatever has focus right now

Window fields: id (the KWin handle — pass it back as window=), name, app_id, resource_name, desktop_file, role, icon, x, y, width, height, minimized, fullscreen, maximized, keep_above/keep_below, no_border, skip_taskbar/pager/switcher, window_type, layer, desktops, plus geometry/position/visible.

How the handles are found (and what they can and cannot do):

• KWin's org.kde.KWin.ScreenShot2 CaptureWindow(handle) takes a window's internal id (a QUuid). There is no list method on that interface, so the helper asks KWin's krunner interface (/WindowsRunner, empty query → every window) for the ids and org.kde.KWin getWindowInfo(handle) for the details. Neither needs the desktop-file authorisation; only the pixel grab does.
• Normal application windows only. Panels/docks, the desktop/wallpaper and overlay windows are not in list_windows() (that is KWin's own filter). If you get hold of one of their handles anyway, Capture(window=…) captures it.
• decoration=False (default) grabs the client area; decoration=True grabs the window with its title bar and shadow — the shadow area is transparent (alpha 0), so composite it or drop the alpha channel before saving a JPEG.
• A window that is covered by other windows is not clipped: KWin renders that window's own buffer, so you get the whole client area, and nothing of the windows on top. A minimised window returns a frame, but it is a stale snapshot: KWin stops rendering windows that are minimised, so you get the last buffer it held — the same bytes every time. Capture(stale_check=True) makes grab() consult KWin (throttled) and set stale_frame so this cannot pass unnoticed, and warn once when a window goes stale; Capture.window_minimized asks on demand. Capture the window while it is mapped if you need its current contents.
• A handle is only valid while the window lives. If the window is closed, grab()/latest() raise WindowGone — the helper stays alive, the other captures are unaffected, and list_windows() no longer reports it.
• Names are resolved by ranking: exact caption > exact caption/app id (any case) > unique substring. Several equally-good matches raise AmbiguousWindow (with .candidates), none raises WindowNotFound (the message lists what is available).

fps, 2560×1440 Plasma 6.6	median
whole screen	26 ms (≈40 fps)
692×440 window	4.7 ms (≈185 fps)

A window grab is cheap because the compositor only has to render and ship that window — fine for watching a handful of windows in a loop.

Monitors: list them, capture one by name or id

kwcapture monitors                    # index, id, name, make/model, size, scale, position
kwcapture monitors --json             # machine-readable
kwcapture monitors --measure-scale    # also probe the real scale (125% / 150% displays)
kwcapture grab --monitor HDMI-A-1 -o tv.png
kwcapture grab --monitor 0 -o left.png

import kwcapture as K

for m in K.list_monitors():
    print(m.index, m.id, m.name, m.position, m.geometry, m.refresh_hz, m.scale)
# 0 65 DP-1     (0, 0)    (2560, 1440) 164.69 1
# 1 79 HDMI-A-1 (2560, 0) (3840, 2160)  60.0  1

Three ways to name the same screen — list_monitors() returns them as Monitor objects, and Capture(monitor=...) takes any of them (or a Monitor):

you pass	what it is
"DP-1"	the connector name — what KWin addresses screens by. Substrings work ("hdmi")
65	the compositor's own id for the output (its wl_output global name)
0	the list_monitors() index, top-left monitor first

K.Capture(monitor="HDMI-A-1")     # by name
K.Capture(monitor=79)             # by compositor id
K.Capture(monitor=1)              # by index
K.Capture(monitor=K.find_monitor("lg tv"))    # by make/model, via the finder

An id is stable for the session and unique; a name is stable across sessions and readable; an index shifts if you plug a monitor in on the left. find_monitor() raises MonitorNotFound listing what exists, or AmbiguousMonitor (with .candidates) when a substring matches several — it never picks one silently. Because the name is resolved eagerly, a typo raises immediately instead of failing deep inside the compositor.

With monitor= set, area= is relative to that monitor, which is usually what you want on a mixed-DPI desk:

K.Capture(monitor="HDMI-A-1", area=(0, 0, 800, 600))   # top-left 800x600 of the TV
K.Capture(area=(0, 0, 800, 600))                       # unchanged: global scene coords

Only enabled outputs are listed — a connected-but-disabled monitor has no image to capture (kscreen-doctor -o shows those and switches them on).

Fractional scaling (75 %, 125 %, 150 %, ...)

A whole-output frame is always device resolution — every pixel the panel shows, never a rescale. What scaling changes is that the numbers you see in the UI are no longer numbers in the image, and there is more than one scale involved. On a desk measured live (DP-1 at 75 %, HDMI-A-1 at 125 %):

what	where you see it	this desk
integer hint	Monitor.scale (wl_output)	1 and 2 — cannot express fractions
display scale	Monitor.effective_scale, Capture.scale of a whole-output frame	0.75 and 1.25
scene factor	Monitor.area_scale, Capture.area_scale, Capture.scale of a region frame	1.25 on both

device pixels = logical pixels x display scale — that is the relation that governs output geometry, window geometry and whole-output frames. kwcapture measures it (grabbing one small area in device pixels and once composited, and reading a whole-output frame), because wl_output can only advertise an integer.

m = K.list_monitors(measure_scale=True)[0]
m.scale             # 1     <- what wl_output advertises (integer hint)
m.effective_scale   # 0.75  <- what the display is actually running
m.area_scale        # 1.25  <- the scene-wide factor region captures come back at
m.fractional        # True
m.logical_geometry  # device size / display scale
m.to_physical(100, 60)    # logical -> device pixels
m.to_logical(150, 90)     # device pixels -> logical

cap = K.Capture(monitor=m)
cap.scale        # scale KWin applied to these frames (cheap, straight from the frame)
cap.pixel_scale  # this output's measured display scale (cached after the first probe)

Region captures are the exception. area= is in logical scene coordinates, and KWin answers a region at the scene factor — one number for the whole desktop, not the scale of the output the region is on. So area=(0, 0, w, h) hands you a w*area_scale x h*area_scale image, and where area_scale != pixel_scale those pixels are a resample of the w*pixel_scale device pixels the region really covers (on a 75 % output, a region grab is an upsample; on an output at the scene scale it is 1:1). If that bothers you — and for pixel-exact work it should — grab the monitor and slice the array instead, below.

area_in="physical" means the rectangle is in device pixels of that output (it needs monitor=, since device pixels only mean something within one output). It is converted through the display scale, so you get exactly the region you named — while the image itself stays at the scene factor, i.e. w * area_scale/pixel_scale px wide:

K.Capture(monitor="DP-1", area=(0, 0, 640, 360), area_in="physical")

For cropping, the simplest exact route is to grab the whole monitor and slice the numpy array — no rounding, no coordinate maths:

frame = K.Capture(monitor="DP-1").grab()
crop = frame[100:460, 640:1280]      # device pixels, exactly

For cropping, the simplest exact route is to grab the whole monitor and slice the numpy array — no rounding, no coordinate maths:

frame = K.Capture(monitor="DP-1").grab()
crop = frame[100:460, 640:1280]      # device pixels, exactly

Resilience: resolution changes, window resizes and a dead daemon

Long-running captures outlive their setup. Unplugging a monitor, switching mode, or maximising the window you are watching all change the frame size, and the helper is a separate process that can be killed, crash, or quit after sitting idle. By default grab() heals all of these by itself — you keep getting frames:

import kwcapture as K

cap = K.Capture(window="Dolphin")     # auto_restart=True is the default

while True:
    frame = cap.grab()               # survives a restart, a mode change, a resize
    if cap.resized:                  # the frame we just got differs in size from the last
        print("now", cap.last_geometry)   # e.g. (2560, 1440) -> (1920, 1080)

• Window resized / maximised — frames simply come back at the new size; there is no need to re-create the Capture. resized is True on the frame where it changed, and last_geometry is the size of the frame you just got.
• Monitor mode change — same thing for a screen capture: geometry() always reports what the compositor is capturing now, and the ring is sized for it.
• A frame bigger than the ring (a bigger monitor, a window grown past the headroom) raises RingTooSmall. The ring cannot be grown in place — a client has it mapped at the old length — so the helper reports it and exits, and a new one is started with a ring sized for the new geometry. grab() does that for you; watch auto_restarts and last_restart_reason if you care that it happened.
• Daemon killed, crashed, or reaped for being idle — restarted transparently and the frame retaken. An outstanding zero-copy view from before the restart stays readable (frozen at its last frame) rather than turning into invalid memory.

Knobs, all on Capture:

parameter	default	meaning
auto_restart	True	recover inside grab(); False restores the old "raise DaemonDead" behaviour
restart_limit	5	give up after this many consecutive restarts (resets on the next good frame)
slot_floor	(5120, 2880)	room left in each ring slot for a frame that grows later; raise it for 6K/8K displays

What is deliberately not auto-recovered, because restarting would hide a real bug:

• WindowGone — the window you were capturing was closed. Restarting cannot bring your window back. (If the app is reopened under the same name, the Capture follows it to the new window handle.)
• TimeoutError while the daemon is alive — KWin is being slow or is wedged. Spawning another daemon would just hide that.

cap = K.Capture(auto_restart=False)   # old behaviour: report it, do nothing about it
try:
    cap.grab()
except K.RingTooSmall:
    cap.restart()                     # re-open the ring at the current geometry
except K.DaemonDead:
    ...                               # you own the restart

Performance notes (2560×1440@165 Hz, Plasma 6.6, i9-13900K)

• ~11 ms is a fixed cost inside KWin (a 320×180 grab still takes 11 ms); 1440p adds ~7 ms of grab and ~5 ms to move 14 MB across the pipe.
• ~47 fps is the hard ceiling — KWin serialises screenshot jobs. depth=2 reaches it; depth=3/4 only add latency (37/59/81 ms per frame).
• Post-processing: OpenCV INTER_AREA 2560→1280 ≈ 0.1 ms, cvtColor ≈3 ms, JPEG 1280 ≈3.5 ms / 2560 ≈4.5 ms, PNG ≈23 ms (prefer JPEG), Pillow resize ≈17 ms.
• Without OpenCV or Pillow, downscaling falls back to numpy (box-average on integer factors, nearest otherwise) and image encoding raises a hint to install an extra.
• The ring file looks like 226 MB but is sparse: 15 MB resident at 1440p.

Repo layout

kwcapture/
  __init__.py            Python API (Capture, Window, grab/shot, shm ctypes mirror)
  __main__.py            CLI: doctor / setup / install-desktop / screens / windows /
                         grab / demo / bench
  _native.py             find or compile the helper ($KWCAPTURE_BIN, wheel, cache, source)
  _desktop.py            the KWin authorisation desktop entry
  native/kwcapture.c     the capture daemon (sd-bus + shm ring + output and window listing)
  native/include/kwcapture_shm.h   shared-memory protocol (asserted on both sides)
tests/test_kwcapture.py  159 functional checks (also `pytest tests/`)
bench.py                 comparison against mss and PIL.ImageGrab
probe/                   the reverse-engineering experiments (Wayland global dumper, etc.)
AGENTS.md                investigation log — how the KWin API and its auth really work

Development:

python3 -m venv .venv && .venv/bin/pip install numpy pillow opencv-python-headless
make setup          # build the helper into kwcapture/bin + authorise
make test bench     # tests, then the comparison table
make wheel          # dist/*.whl

Troubleshooting

symptom	cause / fix
The process is not authorized to take a screenshot	kwcapture install-desktop, then retry (KDE's service cache notices the new file asynchronously — Capture already retries for a few seconds)
no frame within 2.0s	helper died: Capture.restart(), or Capture(daemon_stderr=sys.stderr) to see its log; kwcapture doctor
failed to compile the kwcapture helper	install build-essential libsystemd-dev libwayland-dev, or build it yourself and set KWCAPTURE_BIN
frame needs N bytes, slot has M	resolution went above ~5K: restart the helper
no window matches 'x'	kwcapture windows for the handles; the error lists what is there. Matching is exact caption/app id, or a unique substring
AmbiguousWindow	several windows share that name/caption — pass the handle from kwcapture windows
WindowGone: the window has nothing to capture	the window was closed (or is being unmapped). The helper is fine: list_windows() again and make a new Capture
unsupported kwcapture ABI 1	an old helper binary ($KWCAPTURE_BIN, or a stale kwcapture/bin): rebuild with kwcapture setup
frame N reports an invalid geometry / outside the …-byte ring	the client refused a frame descriptor that cannot fit the ring (old/mismatched helper, or a ring truncated under it). Capture.restart(); if it repeats, kwcapture setup to rebuild the helper
the ring was replaced by restart()/close() while this frame was being read	you grabbed from another thread while something called restart(); a single Capture is not thread-safe for lifecycle + reads
works in a terminal but not from cron/SSH	you need WAYLAND_DISPLAY and DBUS_SESSION_BUS_ADDRESS of the graphical session
colours wrong somewhere	grab() is BGRA; shot()/to_rgb() are RGB

Non-KDE compositors need ext-image-copy-capture-v1 / wlr-screencopy-unstable-v1 instead (kwcapture does not implement those; probe/globals.c shows how to check what a compositor advertises).

License

MIT — see LICENSE. AGENTS.md documents the reverse engineering behind it.