Kwin MCP README

kwin-mcp

Model Context Protocol server for Linux desktop GUI automation on KDE Plasma 6 Wayland

A Model Context Protocol (MCP) server that enables AI agents (Claude Code, Cursor, and other MCP clients) to launch, interact with, and observe any Wayland application in an isolated virtual KWin session -- without affecting the user's desktop display or input. It also supports live desktop automation by connecting to an existing KWin session (real desktop or container) for collaborative workflows. With 29 MCP tools covering sessions, apps, screen capture, the AT-SPI2 accessibility tree, window management, mouse, keyboard, touch, clipboard, and D-Bus, kwin-mcp provides everything needed for end-to-end GUI testing and desktop automation on Linux.

[!IMPORTANT] kwin-mcp 1.0 renamed and reshaped the tools. The 29 tools of 1.0 replace the 37 tools of 0.12, with shared argument types and structured output on every tool. A removed tool name returns an error that names its replacement. See Migrating from 0.11 (also covers 0.12).

Table of Contents

• Why kwin-mcp?
• Use Cases
• Quick Start
• Configuration
• Server Flags
• Available Tools
• Shared Argument Types
• How It Works
• System Requirements
• Installation
• Migrating from 0.11
• Limitations
• End-to-End Testing
• Contributing
• License

Why kwin-mcp?

• Isolated sessions -- Each session runs in its own dbus-run-session + kwin_wayland --virtual compositor. Your host desktop never receives its windows or input. This is not a filesystem or network sandbox; see Isolation scope.
• Live session support -- Connect to a real KDE Plasma desktop or a KWin instance inside a container (e.g. systemd-nspawn) for collaborative "share my screen" workflows.
• No screenshots required for interaction -- The AT-SPI2 accessibility tree gives the AI agent structured widget data (roles, names, coordinates, states, available actions). Every element comes with a ref that any pointer or touch tool accepts as a target, and element_act invokes an element's action without pointer coordinates at all.
• Typed results -- Every tool declares an MCP outputSchema and returns structuredContent next to a readable text rendering, so an agent can read window ids, element rectangles, and frame paths without parsing text.
• Zero authorization prompts -- Uses KWin's private EIS (Emulated Input Server) D-Bus interface directly, bypassing the XDG RemoteDesktop portal. No user confirmation dialogs.
• Works with any Wayland app -- Anything that runs on KDE Plasma 6 Wayland works: Qt, GTK, Electron, and more. Input is injected via the standard libei protocol.
• Full input coverage -- Mouse, keyboard, multi-touch, and clipboard -- all injected through the isolated session for complete desktop automation.

Use Cases

Automated GUI Testing

Run end-to-end GUI tests for KDE/Qt/GTK applications in headless isolated sessions. kwin-mcp launches each app in its own virtual KWin compositor, interacts via mouse, keyboard, and touch input, then verifies results through screenshots and the accessibility tree -- all without a physical display.

AI-Driven Desktop Automation

Let AI agents like Claude Code autonomously operate desktop applications. The agent reads the accessibility tree to understand the UI, performs actions through 29 MCP tools, and observes the results via screenshots -- creating a complete feedback loop for any Wayland application.

Live Desktop Collaboration

Connect to your real desktop session and let Claude observe and interact with what you see. Use session_connect or pass --default-live-session to make live mode the default. Also supports attaching to KWin running inside containers (e.g. systemd-nspawn) for isolated agent desktops.

Headless GUI Testing in CI/CD

Integrate Linux desktop GUI testing into CI/CD pipelines. kwin-mcp's virtual sessions require no X11 or physical display server, making it suitable for headless environments like GitHub Actions or GitLab CI runners on Linux.

Kiosk and Embedded Device Automation

Automate kiosk interfaces and embedded Linux desktops running KDE Plasma or a bare KWin Wayland compositor. Use session_start for isolated virtual testing of kiosk UIs, or session_connect to attach directly to a live kiosk or embedded device session for real-time automation and diagnostics.

Quick Start

Requires KDE Plasma 6 on Wayland. See System Requirements for details.

1. Install system and build dependencies

uv tool install and uvx always install kwin-mcp into an isolated Python environment, and pip install does the same when run inside a virtual environment. An isolated environment cannot reuse your distribution's python3-gi or python3-dbus packages, so PyGObject, pycairo, and dbus-python are built from source during installation and need a C compiler and development headers. Install the packages from Installing System Dependencies first. On Debian 13 (Trixie):

sudo apt-get install -y --no-install-recommends build-essential pkg-config python3-dev libcairo2-dev libgirepository-2.0-dev libdbus-1-dev

2. Install (uv itself: see the uv installation guide)

# Using uv (recommended)
uv tool install kwin-mcp

# Or using pip
pip install kwin-mcp

3. Configure Claude Code

Add to your project's .mcp.json:

{
  "mcpServers": {
    "kwin-mcp": {
      "command": "uvx",
      "args": ["kwin-mcp"]
    }
  }
}

4. Use it

Ask Claude Code to launch and interact with any GUI application:

Start a KWin session, launch kcalc, and press the buttons to calculate 2 + 3.

Claude Code will autonomously start an isolated session, launch the app, read the accessibility tree to find buttons, click them, and take a screenshot to verify the result.

Configuration

Recommended: install as a plugin

The fastest way to wire kwin-mcp into your editor is to install one of the bundled plugins. Each plugin auto-registers the MCP server and ships the kwin-desktop-automation skill, which teaches the agent which tool to call when (session-mode selection, the observe → act → verify loop, element refs, US-QWERTY vs Unicode typing, global logical coordinates, held input, and other platform pitfalls).

Claude Code — install the plugin from the marketplace:

/plugin marketplace add isac322/kwin-mcp
/plugin install kwin-mcp@kwin-mcp

OpenCode — add the npm plugin to the plugins array in your opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": ["@isac322/kwin-mcp-opencode"]
}

For manual OpenCode V2 setup without the plugin, add kwin-mcp under mcp.servers:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "servers": {
      "kwin-mcp": {
        "type": "local",
        "command": ["uvx", "kwin-mcp"]
      }
    }
  }
}

The bundled skill's source of truth is integrations/claude-code/skills/kwin-desktop-automation/SKILL.md. To customize or override it, place a SKILL.md in one of OpenCode V2's skill locations:

• Personal: ~/.claude/skills/kwin-desktop-automation/SKILL.md, ~/.config/opencode/skills/kwin-desktop-automation/SKILL.md, or ~/.agents/skills/kwin-desktop-automation/SKILL.md
• Project: <project>/.claude/skills/kwin-desktop-automation/SKILL.md, <project>/.opencode/skills/kwin-desktop-automation/SKILL.md, or <project>/.agents/skills/kwin-desktop-automation/SKILL.md

Claude Code

Add to your project's .mcp.json:

{
  "mcpServers": {
    "kwin-mcp": {
      "command": "uvx",
      "args": ["kwin-mcp"]
    }
  }
}

Or if installed globally:

{
  "mcpServers": {
    "kwin-mcp": {
      "command": "kwin-mcp"
    }
  }
}

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "kwin-mcp": {
      "command": "uvx",
      "args": ["kwin-mcp"]
    }
  }
}

Running Directly

# As an installed script
kwin-mcp

# As a Python module
python -m kwin_mcp

# Print the server's usage and options, then exit (does not start the server)
kwin-mcp --help

# Interactive CLI (REPL for rapid testing)
kwin-mcp-cli

# Live session mode (default to real desktop instead of virtual)
kwin-mcp --default-live-session
kwin-mcp-cli --default-live-session

# Attach captured frames to tool results as MCP image content
kwin-mcp --screenshot-images

# Downscale every captured frame so its longer side is at most 2576 pixels
kwin-mcp --screenshot-images --screenshot-max-edge 2576

# Never ask before closing a window on the live desktop; refuse instead
kwin-mcp --live-consent refuse

kwin-mcp-cli exposes the same 29 tools as commands with the same names and parameters, and prints the same text the MCP server returns. Arguments are key=value pairs; object, list and union parameters take JSON:

kwin-mcp> session_start
kwin-mcp> app_launch command=kcalc
kwin-mcp> element_query locator='{"app": "kcalc", "role": "push button"}'
kwin-mcp> pointer_click target='{"ref": "e1.12"}'
kwin-mcp> session_stop

Server Flags

The server accepts only --default-live-session, --screenshot-images, --screenshot-max-edge N, --live-consent {elicit,refuse}, and -h/--help. -h/--help anywhere on the command line prints the usage and exits 0, whatever else is present. An unknown option, an abbreviation such as --default, a positional argument, a --screenshot-max-edge value that is not ASCII digits, a --live-consent value other than elicit or refuse, a flag without its value, or a repeated --screenshot-max-edge or --live-consent is a usage error: the server prints the usage and kwin-mcp: error: ... to stderr and exits with status 2 without starting.

Screenshot Images in Tool Results (--screenshot-images)

By default, screen_capture and the capture frame bursts of the action tools return only the PNG file paths (and kwin://screenshots/{name} resource URIs), so the agent must read the files to look at them. Start the server with --screenshot-images to also attach each captured PNG to the tool result as MCP image content: the single screen_capture frame and every burst frame of the 15 tools that take capture, in capture order. The text of the result and its structured content are unchanged; image blocks are only appended after the text, and only for calls that actually captured frames.

Claude Code:

claude mcp add kwin-mcp -- uvx kwin-mcp --screenshot-images

Any JSON MCP config takes the flag as another argument:

{
  "mcpServers": {
    "kwin-mcp": {
      "command": "uvx",
      "args": ["kwin-mcp", "--screenshot-images"]
    }
  }
}

The flag is off by default because every attached frame is sent to the model and costs context tokens, and a capture burst attaches all of its frames — there is no cap. In a live session (session_connect or --default-live-session) the images are your real desktop's pixels, so enabling the flag sends them to the model provider.

Screenshot Size (--screenshot-max-edge, region, max_edge)

A screenshot of a large desktop can be too big for a model. A 7680x1440 workspace is an 11-megapixel PNG of several megabytes. A model provider downscales an oversized image before the model sees it, and the model's pixel coordinates then refer to the resized image, which the Coordinate space line does not describe. kwin-mcp can bound and crop the image itself, so the line stays exact:

• --screenshot-max-edge N downscales every saved screenshot and capture frame so its longer side is at most N pixels, keeping the aspect ratio. 0, the default, keeps full resolution. The bound must keep the image within the provider's limits, or the provider resizes it again and the formula below no longer matches the model's coordinates. According to Anthropic's vision documentation (checked 2026-10-10), Claude 4.7 and later models take at most 2576 pixels on the long edge and 4784 visual tokens of 28x28 pixels; earlier models take 1568 pixels and 1568 tokens. 2576 suits a wide desktop such as 7680x1440 (2576x483); a 16:9 image fits at 2576x1449 on the newer models and 1456x819 on earlier ones, and a squarer image or crop needs a smaller bound. N must be ASCII digits, and the flag may be given only once; anything else is a usage error (exit status 2) before the server starts.
• screen_capture(max_edge) and capture={max_edge} override the flag for one call; max_edge=0 returns full resolution.
• screen_capture(region={x, y, width, height}) (global logical coordinates) crops the capture to that rectangle; capture={region} does the same for a burst. The part outside the captured workspace is dropped, and the Coordinate space line reports the crop's origin and size. A region that misses the workspace entirely is an error.

When an image was downscaled, its Coordinate space line keeps the logical origin and size and adds the image's pixel size and the mapping from the saved image's pixels: each image pixel maps to the logical point at its source pixel's center, for example image 2576x483 downscaled: pixel (px, py) shows (0 + (px + 0.5) * 7680 / 2576 - 0.5, 0 + (py + 0.5) * 1440 / 483 - 0.5). The structured Frame.mapping carries the same scale and origin. Small text is unreadable in a downscaled overview of a wide desktop, so take the overview first, then a region around the area of interest at full resolution.

Live Window Close Consent (--live-consent)

On a live desktop, window_close could discard unsaved work in a real application, so kwin-mcp asks the user before it closes anything:

• --live-consent=elicit (the default) shows the user an MCP form-elicitation prompt: kwin-mcp wants to close window <id> (<app>) on your desktop. Unsaved work in that window may be lost., with a "Close this window" checkbox. The window closes only when the user accepts with the box checked. A decline tells the agent not to retry unless the user asks. A cancel (consent: "cancelled") is not treated as a refusal, because a headless client such as claude -p cancels every prompt without showing it to anyone: the result says the confirmation was dismissed or never shown and points the agent to the app's own UI if the user asked for the close. A client that does not support form elicitation gets a non-error refusal (consent: "unsupported") instead of a prompt.
• --live-consent=refuse never prompts and refuses every live window_close with the same non-error refusal, which points the agent to the app's own UI. Use it for headless clients that cannot answer a prompt, such as claude -p: with the default, such a client cancels the prompt, and in our runs Claude then declined to close the window any other way; with refuse it closed the window with the app's shortcut.

Virtual sessions are unaffected: they close without asking (consent: "not_required"). The result reports the outcome in consent (not_required, accepted, not_confirmed, declined, cancelled or unsupported) and, when nothing was closed, refusal (no_window, not_closeable or no_consent). Consent is a courtesy against an agent mistake, not a security boundary; the real boundary is your client's per-tool approval.

Available Tools

kwin-mcp registers 29 tools. Parameters are listed in schema order: a bare name is required, name?=default is optional with its default (JSON). Every coordinate, in and out, is a global logical pixel. The argument types (Target, Locator, WindowSelector, Capture, ...) are described in Shared Argument Types.

Session Management (4 tools)

App Launch and Logs (2 tools)

Screen Capture (1 tool)

Accessibility Elements (2 tools)

Window Management (5 tools)

Mouse Input (5 tools)

Keyboard Input (3 tools)

Touch Input (3 tools)

Clipboard (2 tools)

Escape Hatches (2 tools)

Output, Errors, Annotations, and Progress

Every tool declares an MCP outputSchema and returns structuredContent that matches it, plus a text block with the same data rendered for reading; kwin-mcp-cli prints that same text. Clients that read only the text lose nothing. Action results carry at (the point that was used), element (when the target was a ref), frames, held (the buttons and keys still held) and warnings; the text ends with Still held: ... whenever something is held.

A call that does not do what was asked comes back as an MCP tool error (isError: true, text Error executing tool <name>: <message>), and kwin-mcp-cli prints it as one Error: <message> line on stderr. The message has two lines: what failed, then a remedy the agent can copy, for example No session running. / Call session_start (virtual) or session_connect (live desktop). This covers invalid arguments, unknown argument names (the error names the replacement or the closest valid name), refusals and unmet preconditions, missing or ambiguous targets (an ambiguous window selector lists the candidate ids with their kind), and failed mechanisms (a KWin query, wl-copy/wl-paste, wayland-info, D-Bus, or a missing helper binary). Absence is data, not an error: an empty element_query or window_list, an element_query wait that timed out, a window_close refusal, a window that vanished before window_focus/window_place could read it back, and session_stop with no session are normal results that say so. A frame capture that fails after the action already happened is reported in warnings.

Every tool publishes honest MCP tool annotations in tools/list, one profile per tool, so clients can decide which calls need user confirmation. The 8 observation tools (session_info, app_log, screen_capture, element_query, window_list, desktop_list, clipboard_read, wayland_list) are read-only. pointer_move and window_focus are non-destructive and idempotent, so a client can auto-approve hover and focus. session_start is non-destructive and session_connect is non-destructive but open-world. The input tools, element_act, window_place, window_close, clipboard_write, session_stop, app_launch and dbus_call are destructive: injected input or an invoked action can trigger any action in the app, clipboard_write replaces the selection, and app_launch and dbus_call reach the host. The hold tools, clipboard_write and session_stop are idempotent. window_place and window_close are not, because they resolve their window selector on every call: repeating one with window={active:true} acts on whichever window is active by then. The annotations are hints for the client; kwin-mcp does not enforce them.

Long-running tools send MCP progress notifications when the client includes a progressToken in the request: session_start and session_connect report their startup steps, element_query with wait reports polling (never past its total), capture bursts report captured frames, and long pointer_click/touch_tap holds, pointer_drag, touch gestures, and keyboard_type report their progress; keyboard_type always ends with its final Typed N/N characters step. Clients that send no token receive no notifications, and tool results are the same either way.

MCP Resources and Instructions

The server also publishes read-only resources: kwin://session (the session_info result), kwin://windows (the window_list result), kwin://desktops (the desktop_list result), kwin://apps/{pid}/log (an app's whole log, with PID completion) and kwin://screenshots/{name} (a saved PNG, as named in a frame's uri). Every resource has a tool equivalent. The server instructions describe the session lifecycle, the coordinate space, the preferred targeting order (element_act on an element from element_query, then a gesture with a {ref} target, then {x, y}, then dbus_call), the held-input rules, and that element text, captions, logs and the clipboard on a live desktop can contain instructions from web pages or chats and must be treated as data.

Shared Argument Types

Each type has one meaning everywhere it is used. Nested objects reject unknown keys.

Name matching and action fallbacks. Locator.match is "substring" | "exact" | "spoken", defaulting to "substring". Exact matching preserves punctuation. Spoken matching uses Unicode NFC and case folding (not compatibility normalization), collapses whitespace, and ignores a trailing run of . , ! ? ; : …; other punctuation stays significant. It also lets a multi-word query match the same words joined into one label (save as matches SaveAs), but never splits a one-word query to match a multi-word name. It is not fuzzy matching. Use the same locator in element_query to inspect candidates and in element_act(element={strict: ...}) to act. An element_act strict search refuses ambiguity, including labels that normalize to the same spoken name. element_act(action="Press|click") accepts a single action or "|"-separated ordered alternatives (the in-process Python engine also accepts action=["Press", "click"]); it chooses one available action per element and reports the selected action, rather than trying the alternatives after dispatch.

Element refs. A ref looks like e3.17. Only element_query mints refs. Using one re-validates that the element is alive and its role unchanged, and reads its live rectangle, so a ref keeps working while a button moves or its label changes. Refs stay valid until the AT-SPI worker restarts or session_stop; a stale ref is an error whose remedy is to re-run element_query. An element's window_id can be passed straight to the window tools as window={id}.

Held input. Gestures (pointer_click, pointer_scroll, pointer_drag, keyboard_press, keyboard_type, touch gestures) release everything they press before they return, even on error. Only pointer_hold(phase="press") and keyboard_hold(phase="press") leave input held. A gesture that would press or release a held control is refused before it sends anything, and the error names the release call; gestures on other controls keep working (typing lowercase text while ctrl is held sends shortcuts, scrolling while left is held drags). If KWin's EIS connection drops, kwin-mcp can no longer send releases: the next result clears the held sets and warns that KWin may keep the input pressed (KWin 6.3.6 and 6.7.5 keep an emulated button down after the disconnect). If KWin replaces an input device (an output or keyboard-layout change), the app never receives a release for that device's controls: they stay held, the next result warns held <controls> stays pressed, and a later release goes through the replacement device. A held key then no longer modifies new keys, because KWin resets its modifier state; the warning says so. Reading the state (session_info) never sends input.

How It Works

Claude Code / AI Agent
       |
       | MCP (stdio)
       v
kwin-mcp server (29 tools)          kwin-mcp-cli (interactive REPL)
  server.py: schemas, gates,          cli.py: key=value / JSON args
  consent, resources                       |
       +--------- both delegate to AutomationEngine (core.py) ---------+
                  models.py: shared argument types + result models,
                  one render() per result (server text == CLI output)
       |
       |-- session_start (virtual) ----> dbus-run-session          (session.py)
       |                                   |-- at-spi-bus-launcher (D-Bus activated)
       |                                   +-- kwin_wayland --virtual
       |                                         +-- [your app]
       |-- session_connect (live) -----> existing KWin (real desktop / container)
       |-- app_launch / app_log -------> subprocess spawn + log file (owned process groups)
       |
       |-- screen_capture / capture ---> KWin ScreenShot2 D-Bus (Spectacle fallback)  (screenshot.py)
       |-- element_query / element_act > AT-SPI2 worker via PyGObject (refs, Action interface)  (accessibility.py)
       |-- window_* / desktop_list ----> KWin scripting (loadScript + D-Bus reply)  (geometry.py)
       |-- pointer_* / keyboard_* /
       |   touch_* --------------------> KWin EIS D-Bus --> libei  (input.py)
       |-- keyboard_type(unicode) -----> wtype, or private Wayland data-control paste  (clipboard.py)
       |-- clipboard_* ----------------> wl-copy / wl-paste (wl-clipboard)
       |-- dbus_call ------------------> dbus-python (in-process, introspected signature)  (dbus_args.py)
       +-- wayland_list ---------------> wayland-info

Triple Isolation (+ Optional Home Isolation)

kwin-mcp provides three layers of isolation from the host desktop:

1. D-Bus isolation -- dbus-run-session creates a private session bus. The isolated session's services (KWin, AT-SPI2, portals) are invisible to the host.
2. Display isolation -- kwin_wayland --virtual creates its own Wayland compositor with a virtual framebuffer. No windows appear on the host display.
3. Input isolation -- Input events are injected through KWin's EIS interface into the isolated compositor only. The host desktop receives no input from kwin-mcp.
4. Home directory isolation (optional) -- session_start(home="isolated") creates a temporary HOME directory with isolated XDG directories (XDG_CONFIG_HOME, XDG_DATA_HOME, XDG_CACHE_HOME, XDG_STATE_HOME). Apps that follow the XDG base directories read and write these instead of host user settings (e.g. ~/.config/kdeglobals), improving test reproducibility. XDG_RUNTIME_DIR is intentionally not isolated because the Wayland socket resides there.

Isolation scope

These layers separate the session's bus, display, and input from the host desktop. They do not confine the processes: launched apps run as your user with your filesystem permissions and network access, and can still reach the original home directory by absolute path when home="isolated". Run untrusted applications in a container or VM rather than relying on a virtual session.

Automatic Session Cleanup at Exit

When the MCP server process ends — on stdio EOF, SIGTERM, SIGHUP, or SIGINT — it calls session_stop, so a session the client never stopped does not outlive the server, and the retention chosen at session start is honored. A virtual session's KWin compositor, its dbus-run-session wrapper bus, and every app launched through app_launch are terminated; in a live session_connect session the user's KWin and pre-existing apps are never signalled — only apps kwin-mcp launched are stopped. Held input is released on both. The exit code follows convention: 0 on a clean EOF, 1 on an unexpected server failure, and 128+signum on a signal (143 for SIGTERM, 129 for SIGHUP, 130 for SIGINT).

The final stop runs on the same single thread as tool calls, serialized with any in-flight tool. If a running tool still holds that thread after a 2 s drain, the server instead terminates the process groups it spawned directly and removes the temporary directories session_stop would remove for the session's retention. kwin-mcp-cli stops the session on SIGTERM/SIGHUP through the same path as Ctrl-C, and on exit also terminates any owned process group that was never attached to a session.

Known limits: SIGKILL runs no cleanup, so with retention.artifacts=false the session's /tmp/kwin-mcp-screenshots-* artifacts directory leaks when the server is killed with SIGKILL (remove it by hand); the next session never reuses it and always creates a fresh directory. A descendant that leaves its launched process group (a double-fork or setsid daemon) escapes; and a process group whose leader was already reaped is left alone, because ownership can no longer be proven.

Input Injection

Mouse, keyboard, and touch events are injected through KWin's private org.kde.KWin.EIS.RemoteDesktop D-Bus interface. This returns a libei file descriptor that allows low-level input emulation without requiring the XDG RemoteDesktop portal (which would show a user authorization dialog). The connection uses:

• Absolute pointer positioning in KWin's global logical coordinates, the space that window frames, element rectangles, and screenshot pixels (after applying the frame mapping) share
• evdev keycodes with full US QWERTY mapping for keyboard input
• Smooth drag interpolation (10+ intermediate steps) for realistic drag operations
• EIS touch emulation for multi-touch gestures (tap, swipe with any number of fingers, pinch)

Screenshot Capture

The normal Wayland capture path tries KWin's org.kde.KWin.ScreenShot2 D-Bus interface first and uses the spectacle CLI as a fallback. Both capture the whole workspace across all outputs. capture bursts on the action tools use the same path, and Pillow converts ScreenShot2's raw pipe frames (RGB32, ARGB32, or RGBX8888 depending on the KWin version) to PNG. The Docker visual QA fixture selects its backend from the installed KWin version: KWin 6.3 and 6.7 use the existing Xvfb/scrot path, while KWin 6.8 and later use a private kwin_wayland --virtual session with /dev/udmabuf and its host group, then compare ScreenShot2 output with an independent Spectacle PNG observer. Normal virtual and live Wayland sessions do not opt into the test-only X11 backend.

Each Spectacle capture has a deadline derived from the image Spectacle must produce, the workspace canvas at the output scale (on mixed-scale layouts, every output is upscaled to the next whole scale above the largest): 15 seconds plus 1 second per canvas megapixel, capped at 120 seconds. A 1920x1080 output gets 18 seconds; a mixed-scale 1.45 + 1.0 layout (a 6490x2160 canvas) gets 30 seconds. A capture that does not finish in time fails with spectacle timed out after <N>s.

Saved PNGs are normalized to KWin's global logical coordinate space, so a screenshot taken on a fractionally scaled output has one image pixel per logical pixel. The result states the mapping, for example on two outputs where a 1280x1024 screen sits left of a 1920x1080 screen:

Screenshot saved: <path> (<size> KB)
Coordinate space: logical: origin (-1280, 0); size 3200x1080; backend screenshot2; coverage full; topology observed stable before/after capture

origin is the top-left of KWin's virtual screen geometry. It is negative when an output sits left of or above (0, 0). To click image pixel (px, py), pass target={x: origin_x + px, y: origin_y + py}; no scale conversion is needed. On a single output at scale 1, the origin is (0, 0) and the size equals the screen size.

Backends deliver different pixel spaces, and kwin-mcp normalizes each one:

• ScreenShot2 CaptureWorkspace already renders logical pixels for the whole virtual screen, including outputs at negative positions.
• Spectacle captures device pixels, which kwin-mcp rescales to logical pixels. When all screens share one scale, Spectacle's composite clips outputs at negative logical positions. kwin-mcp cannot recreate those pixels: it leaves them transparent and reports coverage partial with the captured regions. With mixed scales, Spectacle releases before 6.7.90 (including all Gear-numbered releases such as 24.12) place each screen through OpenCV and abort (or distort the screen) when one lies outside their canvas, for example at a negative position. kwin-mcp detects that layout from the topology and does not start Spectacle; the capture fails with an error naming the output instead.
• X11/scrot (test-only) captures the X root, where each KWin output is an X window of logical size * scale pixels. kwin-mcp locates those windows with xwininfo and rescales each one to logical pixels.

kwin-mcp reads the output topology through KWin scripting immediately before and after each capture and checks the captured image against it, retrying once when the layout changed mid-capture. If the mapping still cannot be proven, screen_capture fails with an error containing coordinate mapping cannot be proven rather than returning guessed coordinates; a burst frame keeps its raw PNG and reports Coordinate space: unavailable (reason) with mapping: null instead. Observed stability is not an atomicity guarantee: a topology change that reverted between the two observations can go unnoticed.

Accessibility Tree

The AT-SPI2 accessibility bus of the session is queried via PyGObject (gi.repository.Atspi) in a worker process. element_query returns a structured list of UI widgets with their roles (button, text field, menu item, etc.), names, states (focused, enabled, visible, etc.), screen coordinates, available actions (click, toggle, etc.), and a ref for each one. Its text uses the tree format (depth indentation, role, name, states, @ screen (x, y, wxh) or @ unavailable (reason)): an empty locator starts with # Accessibility Tree (N elements), a filtered query with Found N elements matching <filters>:.

The virtual session's accessibility bus normally starts through D-Bus activation of the distro's org.a11y.Bus service. Fedora's SELinux policy can deny that activation from a private session bus (fedora-selinux/selinux-policy#3385). Direct startup is disabled by default: only the exact value 1 in the server's KWIN_MCP_ATSPI_DIRECT_START environment variable enables it, not an application's env argument. Neither the server nor a healthy activation path performs direct-start service-file lookup. Only after activation fails does the private session wrapper start an owned child to look up and execute the launcher, then the registry.

For an SELinux-enforcing virtual-session host where activation is denied, opt in in the server environment:

KWIN_MCP_ATSPI_DIRECT_START=1 uvx kwin-mcp

The child reads the named AT-SPI service files from operator-trusted XDG_DATA_DIRS roots, defaulting to /usr/local/share:/usr/share; those roots need not mirror every root D-Bus activation searches. Before executing the launcher, it validates both service commands. The registry child reads and validates its own file again when it starts; if that file has become invalid, direct startup rolls back. Exec= parsing follows dbus-daemon's argument splitting. Acquisition retains the 128 KiB content limit and file-descriptor checks: missing or nonregular candidates and existing invalid-path/link cases remain skippable, while malformed, unreadable, or oversized regular files and inaccessible metadata, paths, or roots stop lookup.

Failed direct startup reports a Warning: line and lets KWin continue. Rollback targets the wrapper's owned Bash jobs, with a shared 3-second TERM grace followed by a shared 2-second KILL grace for the launcher and registry, not a separate wait for each child. Running lookup in those children bounds the caller's wait even when regular-file metadata or reads stall on NFS/FUSE; it does not guarantee that a kernel process stuck in uninterruptible I/O disappears immediately. Launcher descendants may remain in the owned session group until session_stop; programs that escape that group are not sandboxed by this mechanism. Normal D-Bus activation and the existing 60-second session-start watchdog are unchanged: a stall in normal activation itself still relies on that watchdog.

The private helper uses the server's existing sys.executable with -I -S to run the installed, standard-library-only dbus_service_file.py, then os.execv on the parsed executable. It is internal startup machinery, not another installation step or MCP tool. This avoids implicit PATH lookup, shell builtins, function evaluation, and ENOEXEC shell fallback; explicitly configured /bin/sh -c commands and legitimate #! scripts remain supported. Direct processes keep the same user, private bus, session process group, environment, and inherited working directory; relative executables resolve against the server's working directory, not an isolated directory. They start from the server process's SELinux domain rather than through normal D-Bus service activation, which can mean less confinement. The opt-in does not change SELinux policy or provide a general sandbox. Successful direct startup also reports a Warning: line. Tests simulate activation denial; they do not verify this path on an SELinux-enforcing Fedora host.

Firefox, like other apps that check it, exposes an accessibility tree only while the session's org.a11y.Status.IsEnabled switch is on, and a stock Plasma session leaves it off. Virtual sessions switch it on before any app starts; their accessibility bus is private, so the host desktop is unaffected. A live session (session_connect) does not change it by default; the result's accessibility field says whether it is on. Pass enable_accessibility=true to switch it on for the whole desktop until session_stop, which switches it back off (unless it was already on before the connection). The restore turns the switch back off whenever this connection turned it on, so a change the user makes to the switch while connected is overwritten at session_stop; enabling accessibility before connecting keeps it on. Every switch read and write is bounded to a few seconds, so a stalled desktop bus cannot hold session_stop or the server's exit. Firefox follows the change at runtime; an app that reads the switch only at startup must be restarted. The server's exit switches it back off too, including when a tool is still running. Restoring is best-effort: a server killed with SIGKILL, or a bus that refuses the call, leaves the switch on. Launch Chromium with --force-renderer-accessibility to expose web page content: in Chromium 154 the switch alone gave only empty window frames, and the flag gave the page tree whatever the switch.

System Requirements

Optional dependencies:

Installing System Dependencies

kwin-mcp needs two groups of system packages: runtime packages (KWin, Spectacle, AT-SPI2) and build prerequisites for the Python packages it installs from PyPI.

Build Prerequisites for uv and pip Installs

uv tool install kwin-mcp and uvx kwin-mcp always use an isolated Python environment, and pip install kwin-mcp does too when run inside a virtual environment. An isolated environment cannot see the system gi (PyGObject) or dbus modules installed by your distribution, so uv or pip builds PyGObject, pycairo, and dbus-python from source. These builds need a C compiler, pkg-config, the Python headers, and the cairo, GObject Introspection, and libdbus development files. Without them, installation fails while building those packages.

On Debian 13 (Trixie):

sudo apt-get install -y --no-install-recommends build-essential pkg-config python3-dev libcairo2-dev libgirepository-2.0-dev libdbus-1-dev

For other distributions, follow the "Installing from PyPI with pip" build dependency steps in the PyGObject Getting Started guide, and also install your distribution's libdbus development package (it provides the dbus-1 pkg-config file that dbus-python needs).

Runtime Packages

sudo apt-get install -y --no-install-recommends kwin-wayland kwin-common qt6-wayland kde-spectacle libei1 at-spi2-core gir1.2-atspi-2.0

# Optional: for clipboard and Unicode input
sudo apt-get install -y --no-install-recommends wl-clipboard wtype wayland-utils

sudo pacman -S kwin spectacle at-spi2-core python-gobject dbus-python-common

# Optional: for clipboard and Unicode input
sudo pacman -S wl-clipboard wtype wayland-utils

sudo dnf install kwin-wayland spectacle at-spi2-core python3-gobject dbus-python

# Optional: for clipboard and Unicode input
sudo dnf install wl-clipboard wtype wayland-utils

sudo zypper install kwin6 spectacle at-spi2-core python3-gobject python3-dbus-python

# Optional: for clipboard and Unicode input
sudo zypper install wl-clipboard wtype wayland-utils

sudo apt install kwin-wayland kwin-common qt6-wayland kde-spectacle libei1 at-spi2-core python3-gi gir1.2-atspi-2.0 python3-dbus

# Optional: for clipboard and Unicode input
sudo apt install wl-clipboard wtype wayland-utils

Installation

Install the system and build dependencies before using any method below. The uv installs, pip installs into a virtual environment, and the from-source install build PyGObject, pycairo, and dbus-python from source and fail without them.

Using uv (recommended)

uv tool install kwin-mcp

uv also provides uvx, which the .mcp.json examples above use to run kwin-mcp. If the kwin-mcp command is not found after uv tool install, run uv tool update-shell and restart your shell so uv's tool directory is on PATH.

Using pip

pip install kwin-mcp

From source

git clone https://github.com/isac322/kwin-mcp.git
cd kwin-mcp
uv sync
uv run kwin-mcp

Migrating from 0.11

kwin-mcp 1.0.0 is a breaking release with no aliases. Every 0.11 and 0.12 capability is still available, under one of 29 tool names with shared argument types:

• 9 tool names are kept with new parameters or outputs, 28 are removed, and 20 are new. Calling a removed tool returns Unknown tool: <name>. Removed in kwin-mcp 1.0; use <replacement>., so an agent that still knows the old names is told the new call.
• Unknown argument names are now rejected instead of silently ignored. A removed argument such as the old frame-capture delay list returns an error naming its replacement (capture={delays_ms:[...]}).
• Coordinates are objects (target={x, y}, from, to, center, frame={x, y, width, height}), key combinations are one string ("ctrl+shift"), and every action that changes the screen accepts the same capture option.
• session_start no longer launches an app: call app_launch(command) after it.
• 0.12.0 shipped 37 tools, including two virtual-desktop tools; use desktop_list and window_place(window={id}, desktop). Its exact-name, spoken-name and fallback-action options on the old action tool become Locator(query, match) (match="exact" or "spoken") and element_act(element, action="Press|click"). The same matching modes now work in element_query, including waits.
• The typed active-window output added after 0.12.0 becomes window_list(filter={active:true}), observing every window kind in one query. KWIN_MCP_ATSPI_DIRECT_START=1 remains the same server-environment opt-in for virtual sessions.

The 1.0.0 entry of the CHANGELOG has the full tool migration table (every 0.11 and 0.12 tool and the 1.0 call that replaces it), the parameter migration table, and the 24 intentional behavior changes.

Limitations

• US QWERTY keyboard layout only -- keyboard_type(method="keys") supports US QWERTY only: letters, digits, ASCII punctuation, space, tab, and newline. Text containing any other character (non-ASCII such as Korean, CJK, or accented letters, \r, or other control characters) is rejected before anything is typed, with an error naming the unsupported characters; use keyboard_type(method="unicode") for it. Virtual sessions pin the compositor keymap to US, so the host's keyboard layout settings do not change what is typed; a live session attached with session_connect keeps the desktop's own keymap, so its active layout must be US for keyboard_type and keyboard_press to produce the expected characters.

• Unicode typing can paste through the clipboard -- When wtype is unavailable or unsupported by the session, keyboard_type(method="unicode") briefly puts the text on the clipboard, then restores the previous selection and keeps serving it until another copy replaces it, even after session_stop disconnects a live session. The restore is best-effort, not atomic: Wayland has no compare-and-swap for the selection, so a copy another client makes at the same moment as the restore can be overwritten. Wayland also does not tell the clipboard owner which client read the text, so a clipboard manager that reads it after the paste chord can be taken for the target app, and any client that read it keeps its copy. The x-kde-passwordManagerHint marker keeps the text out of Klipper's history, but clipboard managers that ignore the marker can still record it.

• Unicode typing into terminals is limited -- Terminals do not paste on Ctrl+V, and one that does not bind it passes it to the program as ^V, which changes the next key (a Return arrives as a literal CR). keyboard_type(method="unicode") therefore checks KWin's active window first and sends Ctrl+Shift+V to terminals known to paste on it (Konsole, Yakuake, GNOME Terminal, GNOME Console, kitty, Alacritty, WezTerm, foot, Ghostty, st, Xfce Terminal, Tilix, Terminator). Window classes recognized as terminals but with no clipboard paste chord (xterm and urxvt, whose Shift+Insert pastes the primary selection) and windows KWin cannot identify get an error without any key being pressed; a terminal whose class is not recognized still gets Ctrl+V. Only Konsole is exercised by the test suite; the other chords follow each terminal's documented default bindings, so a remapped paste shortcut breaks delivery there.

• KDE Plasma 6+ required -- Older KDE versions or other Wayland compositors (GNOME, Sway) are not supported.

• AT-SPI2 availability varies -- Some applications may not fully expose their widget tree via AT-SPI2.

End-to-End Testing

The Docker suite collects every test under tests/e2e against the packaged application, not an editable source checkout. docker/e2e.Dockerfile builds a wheel and installs it into /opt/kwin-mcp-venv with standard Requires-Dist resolution: PyGObject, pycairo, and dbus-python compile from source in a builder-only stage, while mcp, Pillow, and the remaining dependencies resolve fresh from PyPI within the declared ranges. The suite runs both AutomationEngine tests and the installed kwin-mcp console entry point. The MCP tests initialize a real client/server session over stdio JSON-RPC.

Run the complete suite from the repository root:

scripts/run-e2e-docker.sh

The runner builds the image for Docker's native architecture, creates artifacts/e2e/<UTC-timestamp>-<pid>/, runs pytest, and prints the artifact path. Additional arguments pass directly to pytest:

scripts/run-e2e-docker.sh -- tests/e2e/test_installed_package.py -v
scripts/run-e2e-docker.sh -- -k "visual or screenshot" -v

Fast checks that need no KWin run on the host: uv run pytest tests/unit covers the tool schemas against the design (names, parameters, annotations, output models), unknown-argument and removed-tool errors, the text renderers, held-input rules, selector matching, and the documentation gates (this README's tool tables and the CHANGELOG migration table).

Coverage of the Docker suite includes:

• virtual KWin engine tests for session lifecycle, AT-SPI2 observation, element refs, window listing, placement and closing (including ids that contain a quote), EIS pointer/keyboard/touch input and held input, clipboard, cleanup, and error handling;
• failing-session lifecycle regressions that stub kwin_wayland/dbus-run-session on PATH: session_start must fail within its startup deadline and surface the captured session stderr plus stray stdout (including a newline-free partial line), and teardown must reap the entire owned process group even when the session leader was already reaped or a descendant ignores SIGTERM;
• a KWin bus-name readiness regression whose kwin_wayland stub creates the Wayland socket before starting the real compositor: session_start must wait until org.kde.KWin has an owner before it sets up EIS input, and must fail with a clear error when KWin exits first or never takes the name;
• an accessibility-bus check that org.a11y.Bus has an owner as soon as session_start returns, before any app or AT-SPI2 query could activate it, and that a failed activation (a dbus-send stub on PATH) is reported as a Warning: line in the session_start output;
• the org.a11y.Status.IsEnabled switch: on in virtual sessions, and in live sessions reported, left off by default, switched on by enable_accessibility=true, and restored by session_stop and by a server exit while a tool is still running;
• exact input- and output-schema checks for all 29 registered tools, plus installed-server stdio calls through every MCP tool, resource and the live-consent elicitation flow;
• version-aware visual tests that use Xvfb/scrot on KWin 6.3 and 6.7, and private --virtual ScreenShot2 capture with /dev/udmabuf on KWin 6.8 and later;
• KCalc before/after pixel transitions and a deterministic GUI probe for mouse hover, cursor inclusion, animation frame bursts, and CJK text rendered differently from a missing-glyph tofu control;
• screenshot coordinate mapping at output scales 1.0 and 1.45 (the fractional scale is set through kscreen-doctor): a probe button found by pixel color in the screenshot is clicked at origin plus pixel and must activate, and single captures and frame bursts report the logical workspace as their coordinate space;
• artifact retention and failure behavior, environment provenance, installed distribution metadata (version 1.0.0, the Production/Stable classifier), console entry points, and process/socket cleanup.

Pull requests run every E2E test except the ones marked slow; .github/workflows/e2e-slow.yml runs the slow set.

The container uses software rendering. KWin 6.8 and later actual test containers also receive /dev/udmabuf and its host group for the private virtual compositor; quality and collect-only containers remain device-free. The suite needs no --privileged, --cap-add, physical GPU, or /dev/dri passthrough. KWin 6.3 and 6.7 visual tests keep their test-owned nested Xvfb server, so the host does not need an X server. CI runs the suite natively in four images: Debian trixie (docker/e2e.Dockerfile), Fedora 44 (docker/e2e-fedora.Dockerfile), openSUSE Tumbleweed (docker/e2e-opensuse.Dockerfile), and Arch Linux (docker/e2e-arch.Dockerfile, amd64 only). A failure in any of them fails the workflow.

A completed run retains environment.json, junit.xml, pytest.log, nested-KWin/Xvfb/MCP logs, and visual PNG evidence below its artifact directory. Failed runs also collect Docker inspect, container log, and process-list diagnostics.

Exact-virtual ScreenShot2 coverage follows the installed KWin version. KWin 6.3 and 6.7 keep the legacy opt-in probe and Xvfb/scrot visual path. KWin 6.8 and later actual test containers use /dev/udmabuf with its host group and automatically opt into the exact probe, which compares the product PNG with an independent Spectacle PNG at scale 1.

Provider-unavailable checks use the real org.kde.KWin /Plugins lifecycle: unloading the screenshot plugin removes the provider and makes Spectacle fail without a PNG; reloading the plugin must restore the provider. Semantic pixel, cursor, animation, scaling, crop, retention, and coordinate-mapping failure checks remain in scope. The slow physical-X11 cases stay on the Debian slow lane.

Physical multi-output capture and negative-origin layouts remain unverified, as does Spectacle partial coverage on those layouts. See docker/README.md for the process topology, complete test-file inventory, evidence layout, and targeted commands.

Contributing

Contributions are welcome! See CONTRIBUTING.md for development setup, code style guidelines, and the pull request process.

git clone https://github.com/isac322/kwin-mcp.git
cd kwin-mcp
uv sync
uv run ruff check src/
uv run ruff format --check src/
uv run ty check src/

License

MIT