wbb 0.1.3

Render a live website into an off-screen pixel buffer and expose it as a programmable primitive

wbb — WebView Buffer Bridge

Renders a live website into an off-screen shared-memory pixel buffer and exposes that buffer as a composable, scriptable async Python primitive.

[ headless Chrome ] ──CDP screencast──▶ [ BrowserBridge ]
                                               │
                                         JPEG decode
                                         RGBA write
                                               │
                                               ▼
                                   [ FrameBuffer / SHM ]
                                               │
                                    zero-copy numpy view
                                    filter pipeline
                                               │
                                   ┌───────────┴───────────┐
                                   ▼                       ▼
                            [ DisplayClient ]       [ user script ]
                              SDL2 window           any consumer

github repository

Requirements

• Python 3.10+
• Chrome or Chromium installed on the host (override path with CHROME_PATH env var)
• numpy, Pillow, websockets, aiohttp (auto-installed)
• pygame only if you use DisplayClient (pip install wbb[display])
• scipy optional — filters.blur(radius > 1) uses it if present and falls back to a cheap roll-average otherwise

Installation

pip install wbb                   # core
pip install "wbb[display]"        # + pygame for DisplayClient
pip install "wbb[fast-jpeg]"      # + PyTurboJPEG for faster computation.

Quick start

import asyncio
from wbb import BrowserBridge, FrameBuffer

async def main():
    buf = FrameBuffer("my_buf", 1280, 720)
    async with BrowserBridge(buf) as br:
        await br.navigate("https://example.com")
        frame = await buf.next_frame()
        frame.save("screenshot.png")
        del frame  # release the zero-copy view before the buffer closes
    buf.close()
    buf.unlink()

asyncio.run(main())

---

Public API

FrameBuffer

The central primitive. Owns two named shared-memory segments that form a double-buffer; a metadata segment tracks which side is current.

FrameBuffer(name, width, height, *, attach=False)

Method / property	Description
write(rgba)	Write an H×W×4 uint8 array into the inactive buffer, then flip. Returns the new frame_id. Called internally by BrowserBridge.
read()	Return the latest Frame as a zero-copy, read-only view.
async next_frame(timeout=5.0)	Wait up to timeout seconds for a new frame; on timeout, returns the latest frame via read() instead of raising.
async for frame in buf	Yield one Frame per rendered frame, indefinitely (built on next_frame).
close()	Release this process's memory mappings.
unlink()	Destroy the underlying OS-level segments. Only the creating process should call this.
with FrameBuffer(...) as buf	Context manager: close()s on exit, and also unlink()s if this process created the buffer (attach=False).

Lifetime contract — read this before debugging a hang: read() and async iteration return zero-copy views into shared memory. CPython will not unmap memory while any array still holds a live buffer-protocol export on it — the same rule as mmap generally. In practice: drop every Frame (and anything derived from one, e.g. via .crop()) before calling close() — reassign the variable, let it go out of scope, or del frame. If you need the pixel data to outlive the buffer, call frame.copy() first to detach it entirely. close() itself is best-effort: if a view is still outstanding it logs and returns rather than raising, and the segment is freed once that reference is dropped. unlink() always succeeds regardless, since it only removes the name.

Cross-process attach:

# Process A — creates the buffer
buf = FrameBuffer("shared", 1280, 720, attach=False)

# Process B — connects to the same buffer, no BrowserBridge needed
buf = FrameBuffer("shared", 1280, 720, attach=True)
frame = buf.read()

---

Frame

A snapshot of one rendered frame.

Frame(data, width, height, frame_id, timestamp)

Method	Description
data	np.ndarray H×W×4 uint8 RGBA. Read-only view into shared memory.
copy()	Return a writable, detached copy of the array.
crop(x, y, w, h)	Zero-copy sub-region view. Returns a new Frame (same frame_id/timestamp).
save(path, format=None)	Save to file (PNG, JPEG, …); format inferred from extension unless given.
frame_id	Monotonically increasing integer, unique within a FrameBuffer session.
timestamp	time.monotonic() at write time.

---

BrowserBridge

Owns the headless Chrome process and CDP connection.

BrowserBridge(
    buffer,
    *,
    width=1280,
    height=720,
    screencast_quality=80,
    screencast_max_fps=30,
    enable_input=True,
    headless_args=None,
)

Lifecycle

async with BrowserBridge(buf) as br:
    ...
# or manually
br = BrowserBridge(buf)
await br.start()
...
await br.stop()

stop() is a no-op (returns immediately) if start() was never called.

Navigation

await br.navigate(url)                      # does not block on load — see note below
await br.reload(ignore_cache=False)
await br.set_viewport(width, height)
result = await br.eval("document.title", await_promise=False)
found = await br.wait_for_selector("#some-id", timeout=10.0, poll_interval=0.15)

Note: navigate() sends Page.navigate and returns immediately; it does not wait for the load event. Pace multi-step automation with wait_for_selector(), click_element()'s own polling, or an explicit asyncio.sleep(...) between steps — see examples/03_multi_step_automation.py.

Input (requires enable_input=True, which is the default; all of these raise RuntimeError if input was disabled)

await br.click(x, y, button="left")
await br.move(x, y)
await br.scroll(x, y, delta_x=0, delta_y=100)
await br.key("Enter")          # DOM key name; see caveat below
await br.type("hello")         # inserts text via Input.insertText
found = await br.click_element(selector=None, text=None, nth=0,
                                scroll_into_view=False, timeout=5.0,
                                poll_interval=0.15, button="left")

key() is known to be flaky. It dispatches a raw Input.dispatchKeyEvent and Chrome doesn't always treat it as Wikipedia's search box does, for instance — sending "\r" as a key reliably submits a form where "Enter" does not, in practice. If a key event doesn't do what you expect, try the alternative character, or fall back to eval()-driven form submission.

click_element resolves an element by CSS selector (tried first; if it matches anything, the nth match is used directly — no redirect to a clickable ancestor) or, failing that, by case-insensitive text substring match against an element's own direct text nodes, with each text match resolved to its nearest clickable ancestor (a, button, input, select, textarea, [role="button"], [onclick], summary, label) and de-duplicated. It polls every poll_interval seconds up to timeout, then clicks the resolved bounding-box center via the same path as click(). Returns False on timeout rather than raising; raises ValueError if neither selector nor text is given.

Frame access

frame = br.buffer.read()    # latest frame, immediately
async for frame in br.frames():
    ...                     # one Frame per screencast push

Event hooks

br.on("frame",    lambda frame: ...)      # every rendered frame
br.on("navigate", lambda url: ...)        # page navigation
br.on("load",     lambda: ...)            # load event fired
br.on("console",  lambda level, args: ...)
br.on("error",    lambda description: ...)

Callbacks may be plain functions or coroutine functions.

---

DisplayClient

Optional SDL2 window. Requires pygame.

DisplayClient(
    buffer,
    *,
    title="wbb",
    filters=None,           # list of Filter callables
    on_mouse_event=None,    # (event_type, x, y, button) -> Any
    on_key_event=None,      # (event_type, key_name) -> Any
    window_size=None,       # (width, height); defaults to buffer size
)

window_size is useful when a filter changes the frame's output dimensions (e.g. a crop) — set it to match so the SDL surface isn't mismatched against the buffer's native size.

# Blocking
display.run()

# Async task (composable with other coroutines)
task = asyncio.create_task(display.run_async())
display.stop()   # signal shutdown from another coroutine
await task

Input forwarding to the browser:

def on_mouse(kind, x, y, button):
    if kind == "down":
        asyncio.create_task(br.click(x, y))

display = DisplayClient(buf, on_mouse_event=on_mouse)

---

filters module

Every filter is (np.ndarray) -> np.ndarray — plain callables, fully composable. Built-ins return factory functions so parameters are explicit:

from wbb import filters

pipeline = [
    filters.crop(0, 0, 640, 360),    # top-left quadrant
    filters.scale(1280, 720),        # scale back up
    filters.grayscale(),
    filters.blur(radius=2),          # uses scipy if installed, else a cheap fallback
    filters.brightness(+20),
    filters.contrast(1.2),
    filters.colorize(r=1.1, g=0.9, b=0.9),
    filters.flip(horizontal=True),
]

# Combine into one callable
f = filters.chain(*pipeline)
result = f(frame.data)

User-defined filters plug in identically:

def my_filter(frame: np.ndarray) -> np.ndarray:
    return frame.copy()   # or any transformation

pipeline = [filters.grayscale(), my_filter]
display = DisplayClient(buf, filters=pipeline)

---

Scriptability model

A user scenario is a plain .py file that imports wbb and composes its primitives. The library places no constraints on structure or complexity.

Using the CLI script runner

python -m wbb script my_scenario.py --url https://example.com

The script receives pre-initialized objects as module-level names: wbb_buffer, wbb_browser, wbb_url, plus all public symbols (BrowserBridge, FrameBuffer, Frame, DisplayClient, filters).

wbb_browser is not started for you — the script owns its lifecycle, exactly like every BrowserBridge in examples/ (async with BrowserBridge(...) as br: or manual start()/stop()). This avoids launching a Chrome process the script may never touch if it builds its own objects instead, which is equally valid. If the script never starts wbb_browser, the runner's cleanup stop() call is a safe no-op.

If the script defines an async def main(), the runner calls it. Otherwise the script executes at import time.

CLI commands

# Open a URL in a window (requires pygame)
python -m wbb display https://example.com

# Save a screenshot and exit
python -m wbb screenshot https://example.com output.png --wait 2.0

# Run a user scenario script
python -m wbb script path/to/scenario.py --url https://example.com \
    --width 1920 --height 1080

Shared flags (--width, --height, --buffer-name) belong to the top- level parser and go before the subcommand name.

---

Examples

Script	What it demonstrates
initial_test.py	Minimal smoke test: navigate, grab one frame, save it
01_display_with_filter.py	DisplayClient + user-defined vignette filter
01_1_display_with_filter.py	Same, with a normalized crop region and a matching window_size
02_monitor_and_react.py	Pixel diff loop → conditional click + screenshot, with graceful Ctrl-C shutdown
03_multi_step_automation.py	Navigate → type → key → click_element (selector and text fallback) → screenshot pipeline
04_cross_process.py	BrowserBridge in process A, DisplayClient in process B
05_record_to_ffmpeg.py	Headless recording via ffmpeg pipe, real-time frame pacing, no window
06_combined_scenario.py	Display + record + monitor as concurrent async tasks sharing one buffer

---

Architecture notes

No Playwright, Puppeteer, or CEF. wbb drives Chrome directly via the Chrome DevTools Protocol (CDP) WebSocket. This makes it pip-installable everywhere and lets it fit inside a single small package.

Screencast, not polling. Page.startScreencast pushes JPEG frames as they are rendered. There is no polling loop and no screenshot RPC overhead.

Double-buffered shared memory. Two segments alternate; readers always see a complete frame. The metadata segment (active index + frame_id + timestamp) is the only synchronisation point. Cross-process reading requires no locking — see FrameBuffer's lifetime contract above for the one thing that does require care: dropping zero-copy views before close().

Decoupled layers. Render, buffer, and display each run independently. A slow display does not drop render frames; a slow render does not block the display. User tasks run alongside both without any additional glue.

Resource-tracker hygiene. wbb._shm.ShmSegment wraps multiprocessing.shared_memory.SharedMemory and immediately unregisters attached (non-owning) segments from the process's resource tracker, so an attaching/reader process never races the owning process's unlink() on exit (see wbb/_shm.py for the CPython issue this works around).

---

Changelog

V0.1.1

• improved performance and fixed bugs

V0.1.2

• changed header implementation

---

License

MIT

personal notes

pip install "setuptools>=68" wheel "numpy>=1.24" "websockets>=12.0" "Pillow>=10.0" "aiohttp>=3.9" "PyGObject>=3.50" "PyTurboJPEG>=1.7" rm -rf dist/ build/ .egg-info update version in pyproject.toml python -m build twine check dist/ twine upload dist/* git tag v0.1.1 git push origin v0.1.1