vdisplay 0.1.7

Cross-platform virtual display orchestration with virtual and mirror sessions

vdisplay

AI Cost Tracking

• 🤖 LLM usage: $3.6181 (6 commits)
• 👤 Human dev: ~$356 (3.6h @ $100/h, 30min dedup)

Generated on 2026-06-09 using openrouter/qwen/qwen3-coder-next

---

Cross-platform virtual display orchestration API for Python.

One unified API, multiple OS backends with different capabilities. Monitors and windows include an nl field — a natural-language description of what they contain.

CLI, DSL, REST, MCP, and the local vdisplay-agent broker all route through application.executor and shared application services (src/vdisplay/application/).

Quick start

pip install "vdisplay[pillow]"
# or from source (recommended for development):
pip install -e ".[pillow,dev]"
pip install -e "packages/vdisplay-agent[serve]"
pip install -e packages/dsl2vdisplay packages/rest2vdisplay packages/mcp2vdisplay

unset DISPLAY   # optional: auto-resolves host display to :0
vdisplay all
vdisplay monitors
vdisplay windows --apps-only

Documentation

Doc	Description
docs/index.md	Documentation hub — start here
docs/installation.md	System and Python setup
docs/docker-guide.md	Running in Docker
docs/examples.md	All usage examples by environment
docs/troubleshooting.md	Common errors and fixes
docs/agent-broker.md	vdisplay-agent — install-once broker, env vars, REST/MCP
docs/architecture.md	CommandRequest + executor routing (local vs agent)
examples/README.md	Runnable example projects
packages/README.md	Control layer — DSL, MCP, REST, NL, agent

Examples

Example	Mode	Host X11	Run
headless-virtual	virtual	No	cd examples/headless-virtual && docker compose up --build
agent-broker	broker	No	cd examples/agent-broker && ./run.sh
ci-agent	virtual	No	cd examples/ci-agent && docker compose run --rm ci-agent
dev-workspace	dev	No	cd examples/dev-workspace && docker compose run --rm dev
host-mirror	mirror	Yes	cd examples/host-mirror && ./run.sh
host-relay	relay	Yes	cd examples/host-relay && ./run.sh
run_all_examples.sh	mixed	varies	./examples/run_all_examples.sh

Details: docs/examples.md · per-example READMEs in each folder above.

List monitors, windows, and display state

Each monitor and window includes nl — a human-readable summary of what it contains.

# everything — monitors + application windows + adopted (off-screen)
vdisplay all
vdisplay all | jq '{monitors: .monitor_count, windows: .window_count, adopted: .adopted_count}'

# monitors only (xrandr) — name, geometry, rotation, primary, nl
vdisplay monitors
vdisplay monitors | jq '.monitors[] | {name, primary, nl}'

# application windows only — title, app, pid, class, nl
vdisplay windows --apps-only
vdisplay windows --app "Firefox"
vdisplay windows --class jetbrains-toolbox --pid 32977
vdisplay windows --min-width 400 --min-height 300

# platform capabilities + monitors
vdisplay info

# diagnose DISPLAY, socket, monitor count
vdisplay diagnose

# adopted (off-screen) windows — persisted in ~/.cache/vdisplay/
vdisplay relay list

Command	Shows
vdisplay all	monitors + windows + adopted
vdisplay monitors	connected displays only
vdisplay windows	visible application windows only
vdisplay relay list	windows moved off-screen by relay
vdisplay info	platform capabilities
vdisplay diagnose	DISPLAY diagnostics

vdisplay outputs still works but is deprecated — use vdisplay all or vdisplay monitors. vdisplay relay list-windows is deprecated — use vdisplay windows.

Filter windows with --app, --class, --pid, --min-width, --min-height. Use --all (default) to include internal/helper windows; --apps-only excludes them.

Example output

Monitor:

{
  "name": "DP-2",
  "primary": true,
  "width": 4320,
  "height": 7680,
  "rotation": "left",
  "rotation_degrees": 90,
  "monitor_id": 0,
  "nl": "Primary monitor DP-2 (4320×7680, rotated left (90°)). Visible apps: Toolbox."
}

Window:

{
  "window_id": "8388615",
  "app_label": "Toolbox",
  "monitor_id": 0,
  "monitor_name": "DP-2",
  "monitor_index": 0,
  "pid": 32977,
  "nl": "Toolbox application window (880×1326 at (4470,298)) on monitor DP-2, process jetbrains-toolb, class jetbrains-toolbox."
}

Screenshots

Quick reference

Goal	Command
One monitor (host)	vdisplay screenshot -o screen.png --source DP-2
All monitors	vdisplay screenshot --all-monitors --out-dir ./captures
Virtual Xvfb (no portal)	vdisplay virtual screenshot -o screen.png --display :99
Mirror + capture	vdisplay mirror screenshot -o mirror.png --source primary --target DP-1
Describe saved PNG (img2nl)	vdisplay screenshot -o screen.png → JSON field nl
Loop every 1s (after screencast)	see Continuous capture below

With VDISPLAY_AGENT_URL set, screenshots route through the broker (same providers, one runtime).

export VDISPLAY_AGENT_URL=http://127.0.0.1:8765

# single monitor — use output name from: vdisplay monitors
vdisplay screenshot -o screen.png --monitor 1
vdisplay screenshot -o screen.png --source DP-2

# all connected monitors (DP-2.png, DP-1.png, HDMI-1.png, …)
vdisplay screenshot --all-monitors --out-dir ./captures

# isolated Xvfb session (works without portal / on headless CI)
vdisplay virtual screenshot -o screen.png --display :99
vdisplay screenshot -o vd.png --mode virtual --vd-display :99

# mirror: prefer mirror screenshot on Wayland (uses ScreenCast when active)
vdisplay mirror screenshot -o mirror.png --source primary --target DP-1

Wayland host (GNOME) — start here

On GNOME Wayland, DISPLAY=:0 is XWayland — direct X11/DRM capture often yields black PNGs (especially with NVIDIA).

One-time setup — needs two terminals (broker must stay running):

# terminal 1 — leave this running (blocks the shell)
vdisplay agent serve
# same as: vdisplay-agent serve

# terminal 2 — all capture commands go here
vdisplay agent health                    # should return {"status":"ok",...}
vdisplay agent screencast start          # pick monitor in portal dialog (once)
# GNOME: Settings → Privacy → Screen Recording → allow vdisplay-agent / terminal

Common mistake: running vdisplay agent screencast start before vdisplay agent serve gives Set VDISPLAY_AGENT_URL … or start: vdisplay-agent serve. Start the broker first.

VDISPLAY_AGENT_URL is optional when the broker listens on 127.0.0.1:8765 (auto-detect). Explicit export still works: export VDISPLAY_AGENT_URL=http://127.0.0.1:8765

Then capture as often as you like (no new prompts):

vdisplay screenshot -o host.png --source DP-2
vdisplay screenshot -o host.png --source primary
vdisplay mirror screenshot -o mirror.png --source primary --target DP-1
vdisplay relay screenshot -o relay.png --source DP-2

REST equivalent:

curl -X POST http://127.0.0.1:8765/session/screencast/start \
  -H 'content-type: application/json' -d '{"interactive": true}'
vdisplay screenshot -o host.png --source DP-1

Note: vdisplay mirror start -o mirror.png runs xrandr mirror then tries driver-level capture directly. On Wayland use vdisplay mirror screenshot (or vdisplay screenshot) after screencast is active.

Continuous capture (Wayland)

After vdisplay agent screencast start, grab frames every second without portal prompts:

export VDISPLAY_AGENT_URL=http://127.0.0.1:8765
mkdir -p ./captures
i=0
while true; do
  vdisplay screenshot -o "./captures/frame-$(printf '%04d' "$i").png" --source DP-2
  i=$((i + 1))
  sleep 1
done

Python (virtual display — no portal needed):

import time
from pathlib import Path
from vdisplay import VirtualDisplaySession

vd = VirtualDisplaySession.create(display=":99")
vd.start()
try:
    for i in range(60):
        vd.save_screenshot(f"frame-{i:04d}.png")
        time.sleep(1)
finally:
    vd.stop()

See also examples/ci-agent/ (VD_FRAMES=60 in Docker).

Describe a screenshot (img2nl)

By default, vdisplay screenshot enriches JSON with an nl field — a natural-language description of the saved PNG (Polish locale by default):

vdisplay screenshot -o screen.png --source DP-2 | jq '{path, nl, img2nl}'

Requires optional package: pip install img2nl[analyze]. Disable with VDISPLAY_IMG2NL=0 or vdisplay screenshot --no-img2nl.

To describe an existing image file:

from vdisplay.application.services.img2nl_enrich import describe_screenshot_image
print(describe_screenshot_image("screen.png")["text"])

Actions — virtual, mirror, relay

# virtual display (isolated Xvfb)
vdisplay virtual start --width 1920 --height 1080 --display :99
vdisplay virtual launch xterm -hold          # flags after the command name are OK
vdisplay virtual screenshot -o screen.png --display :99

# mirror — needs two monitors; list names first with: vdisplay monitors
vdisplay mirror start --source primary --target DP-1          # xrandr only (no PNG)
vdisplay mirror screenshot -o mirror.png --source primary --target DP-1   # Wayland: screencast first
VD_SOURCE=DP-2 VD_TARGET=HDMI-1 ./examples/host-mirror/run.sh   # Docker; black PNGs on Wayland

# relay — hide window off-screen, restore later (separate CLI calls OK)
# 1) find a window to move
vdisplay windows --apps-only
vdisplay windows --app "JetBrains" | jq '.windows[] | {window_id, app_label, nl}'

# 2) adopt (move off-screen) — match by app, title, class, pid, or window-id
vdisplay relay adopt-window --app "JetBrains"
vdisplay relay adopt-window --title "Firefox"
vdisplay relay adopt-window --class jetbrains-toolbox --pid 32977
vdisplay relay adopt-window --window-id 8388615

# 3) move to another monitor instead of off-screen
vdisplay relay adopt-window --app "Firefox" --target HDMI-1

# 4) list adopted windows (persisted in ~/.cache/vdisplay/)
vdisplay relay list
vdisplay relay list | jq '.adopted[] | {window_id, app_label, title, nl}'

# 5) restore — separate CLI call works (geometry saved on adopt)
vdisplay relay release-window --app "JetBrains"
vdisplay relay release-window --window-id 8388615

# 6) screenshot while relaying (Wayland: agent serve + screencast first)
# terminal 1: vdisplay agent serve
vdisplay agent screencast start
vdisplay relay screenshot -o before.png --source DP-2
vdisplay relay adopt-window --app "JetBrains"
vdisplay relay screenshot -o after-hide.png --source DP-2
vdisplay relay release-window --app "JetBrains"
vdisplay relay screenshot -o after-restore.png --source DP-2

# 7) full host demo (before / after adopt / after release PNGs)
export VDISPLAY_AGENT_URL=http://127.0.0.1:8765
vdisplay agent screencast start
WINDOW_APP=JetBrains ./examples/host-relay/run-host.sh
# output: examples/host-relay/output/before_automation.png, after_adopt.png, after_release.png

Relay tips

• Adopted window positions persist across CLI calls (~/.cache/vdisplay/__vdisplay_stash__-*.json).
• Relay moves windows within the same X11 session — it does not move apps into Xvfb :99.
• On Wayland, window move uses XWayland; screenshots need agent + screencast (Docker ./run.sh often yields black PNGs — use ./run-host.sh instead).
• Filter windows first: vdisplay windows --app "Toolbox" → use --app, --title, or --window-id from JSON.

Natural language and DSL

# NL → DSL → JSON (built into vdisplay)
vdisplay nlp "list monitors on display zero"
vdisplay nlp "show application windows" --dsl-only

# DSL bus (same semantics as CLI query verbs)
dsl2vdisplay -c 'MONITORS DISPLAY :0'
dsl2vdisplay -c 'WINDOWS DISPLAY :0 APPS_ONLY'
dsl2vdisplay -c 'ALL DISPLAY :0'

# NL package (thin wrapper; pip install -e packages/nlp2vdisplay)
nlp2vdisplay to-dsl "list monitors on display zero"
nlp2vdisplay apply "show application windows on display zero"
# shorthand — bare prompt defaults to apply:
nlp2vdisplay "show application windows on display zero"

Legacy DSL verbs still work: OUTPUTS maps to MONITORS, WINDOWS unchanged.

vdisplay-agent broker

Install once on the host. CLI, DSL, REST, and MCP set VDISPLAY_AGENT_URL and route capture, sessions, and discovery through the broker — no per-app portal prompts or direct DRM access from adapters.

pip install -e "packages/vdisplay-agent[serve]"
pip install -e packages/dsl2vdisplay packages/rest2vdisplay packages/mcp2vdisplay

# terminal 1 — broker (localhost)
vdisplay-agent serve
# or: vdisplay agent serve

# terminal 2 — clients
export VDISPLAY_AGENT_URL=http://127.0.0.1:8765
vdisplay agent health
vdisplay monitors                    # 3 monitors in ~50 ms via agent /outputs
dsl2vdisplay -c 'MONITORS'
rest2vdisplay serve --port 8216 --agent-url $VDISPLAY_AGENT_URL
mcp2vdisplay serve                   # tool: vdisplay_agent_status

Env var	Role
VDISPLAY_AGENT_URL	Client → broker URL
VDISPLAY_AGENT_TOKEN	Optional Bearer auth
VDISPLAY_CAPTURE_ALLOW_PORTAL	Opt-in portal screenshot (agent only)

Wayland host capture: start ScreenCast once in the agent (POST /session/screencast/start), then use normal screenshot commands. See docs/agent-broker.md and examples/agent-broker.

Virtual display screenshots work without portal. DRM/fbdev capture requires video group on some setups.

Full reference: docs/agent-broker.md · packages/vdisplay-agent

Control layer equivalents

Intent	CLI	DSL
Full state	vdisplay all	ALL DISPLAY :0
Monitors	vdisplay monitors	MONITORS DISPLAY :0
Windows	vdisplay windows --apps-only	WINDOWS DISPLAY :0
Adopt window	vdisplay relay adopt-window --app X	ADOPT APP X
Screenshot	vdisplay screenshot -o out.png	SCREENSHOT OUT out.png DISPLAY :99
Validate tools	vdisplay diagnose	VALIDATE DISPLAY :0

Modes

Mode	Purpose	Isolation	Screenshot	Window move
virtual	Private Xvfb session for agents	Yes	Yes	No (use launch())
mirror	Duplicate existing display output	No	Yes	N/A
relay	Move window within same X11 session	Partial	Yes (relay screenshot)	Yes
screencast	Portal ScreenCast in agent (Wayland)	No	Yes (after consent)	N/A

Requirements (Linux v0.1)

Component	Used by
Xvfb, xwd, scrot	virtual / X11 capture
xrandr	mirror mode
xdotool	relay + input
python3-dbus, python3-gi	portal ScreenCast (Wayland host)
ffmpeg (PipeWire) or GStreamer pipewiresrc	ScreenCast frame grab
Pillow (optional)	faster PNG encoding

sudo apt install xvfb x11-apps x11-utils xdotool scrot x11-xserver-utils
sudo apt install python3-dbus python3-gi   # Wayland ScreenCast in agent
pip install "vdisplay[pillow]"

Full setup: docs/installation.md

Python API

Sessions (backends)

from vdisplay import VirtualDisplaySession, MirrorSession, WindowRelaySession
from vdisplay.discovery import list_monitors, list_windows

# Inspect monitors and windows with nl descriptions
for monitor in list_monitors():
    print(monitor["nl"])

for window in list_windows(apps_only=True):
    print(window["nl"])

# Virtual isolated display
vd = VirtualDisplaySession.create(width=1920, height=1080)
vd.start()
vd.launch(["xterm"])
vd.save_screenshot("screen.png")
vd.stop()

# Mirror existing desktop (same session, no isolation)
# On GNOME Wayland: start agent screencast before save_screenshot, or use capture_host_to_file()
m = MirrorSession.create(source="primary", target="DP-1")
m.start()
m.save_screenshot("mirror.png")   # needs ScreenCast on Wayland
m.stop()

# Relay window off-screen and restore (persists across CLI calls)
r = WindowRelaySession.create()
r.start()
r.adopt_window(match_app="JetBrains")
r.release_window(match_app="JetBrains")
r.stop()

Application layer (shared by CLI / DSL / REST / MCP / agent)

from vdisplay.application.commands import CommandRequest
from vdisplay.application.executor import execute
from vdisplay.application.services import discovery, capture, session, info

# Single execution entry (routes to agent when VDISPLAY_AGENT_URL is set)
result = execute(CommandRequest.from_dsl({"verb": "MONITORS"}, line="MONITORS"))
print(result.data)

# Direct service use-cases (always in-process)
monitors = discovery.list_monitors(display=":0")
meta = capture.capture_screenshot(output="screen.png", monitor=1)
session.virtual_start(width=1280, height=720, display=":99")
caps = info.platform_info()

Via broker SDK:

from vdisplay.client import AgentClient

client = AgentClient("http://127.0.0.1:8765")
client.outputs()
client.start_screencast(interactive=True)

Window heuristics (testable submodules)

from vdisplay.windows import list_windows_enriched, find_windows, pick_best_window
from vdisplay.windows.rank import dedupe_app_windows
from vdisplay.windows.filter import is_internal_window

Project layout

src/vdisplay/
├── application/
│   ├── commands.py         # CommandRequest / CommandResult / CommandVerb
│   ├── executor.py         # single entry: execute() → local or agent
│   ├── handlers/           # local + agent command handlers
│   └── services/           # discovery, capture, session, info
├── commands/               # CLI registry (set_defaults per subcommand)
├── windows/                # scan → normalize → filter → rank → query
├── capture/
│   ├── providers/          # drm, fbdev, mss, x11, portal (opt-in)
│   └── portal_screencast.py  # persistent ScreenCast (Wayland)
├── backends/               # virtual, mirror, relay
├── client.py               # AgentClient SDK
└── cli.py                  # thin entry: register_all + args.func(args)

packages/
├── dsl2vdisplay/           # grammar + CQRS bus → executor
├── vdisplay-agent/         # localhost broker (privileged runtime)
├── rest2vdisplay/          # HTTP → DSL
├── mcp2vdisplay/           # MCP tools
└── nlp2vdisplay/           # NL → DSL

Control layer (DSL / MCP / REST / NL)

Programmatic interfaces on top of the same application services. All query results include nl on monitors and windows.

Package	Role
dsl2vdisplay	Grammar + CQRS bus (MONITORS, WINDOWS, ADOPT, …)
nlp2vdisplay	Natural language → DSL
uri2vdisplay	vdisplay://cmd/... → DSL
cli2vdisplay	REPL over DSL
mcp2vdisplay	MCP server tools
rest2vdisplay	REST API on port 8216
vdisplay-agent	Local broker — sessions, capture, relay

vdisplay-agent serve --port 8765
export VDISPLAY_AGENT_URL=http://127.0.0.1:8765

pip install -e packages/dsl2vdisplay packages/rest2vdisplay packages/mcp2vdisplay
rest2vdisplay serve --port 8216 --agent-url $VDISPLAY_AGENT_URL
mcp2vdisplay serve
curl -s http://127.0.0.1:8216/health | jq .
curl -s -X POST http://127.0.0.1:8216/v1/dsl -H 'content-type: text/plain' -d 'MONITORS' | jq .

Full reference: packages/README.md

Limitations

• Existing windows on DISPLAY=:0 cannot move into Xvfb :99 — different X servers.
• Use VirtualDisplaySession.launch() for apps on the virtual display.
• Use WindowRelaySession to hide/show windows on the current session.
• mirror controls the same desktop through a duplicated output, not an isolated copy.
• nl on monitors lists apps whose window center falls on that output geometry.
• On GNOME Wayland, Docker X11 forwarding often produces black screenshots — use vdisplay-agent on the host instead.
• Windows/macOS backends are planned; Linux/X11 + Wayland (via agent) supported in v0.1.

Troubleshooting: docs/troubleshooting.md

Development

pip install -e ".[pillow,dev]"
pip install -e "packages/vdisplay-agent[serve]"
pip install -e packages/dsl2vdisplay packages/rest2vdisplay packages/mcp2vdisplay
pytest tests/ -q
./examples/agent-broker/run.sh
./examples/run_all_examples.sh   # where host X11 is available

Architecture: docs/architecture.md

License

Licensed under Apache-2.0.