som-window-mcp
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
README.md not found. View project on GitHub.
Available Tools
5 toolscapture_som
Capture a window with numbered marks on every interactive element.
This is the tool to use when you need to understand *where* things are, not
just *what* they are. The accessibility tree decides the boxes, so marks are
pixel-exact and never hallucinated, and every mark in the legend carries a
real screen rect.
Prefer this over capture_window when you plan to act on the result: answer
with a mark number, or use the rect directly and skip vision entirely.
max_marks caps how many controls are numbered (dense toolbars get crowded);
the highest-value controls are kept - named ones first, then larger ones.
max_side downscales the image before drawing, so badges stay sharp.
Falls back to a plain unmarked capture with an explanation when the window
exposes no UI tree (Chromium/Electron without --force-renderer-accessibility).
Inputs: window (required), max_side, max_marks.
capture_window
Capture one window's pixels, including while it is covered.
Uses PrintWindow, which asks the window to render itself, so the window does
not need to be in the foreground: a window hidden behind your terminal still
captures cleanly and only that window's content is in the image.
Returns an image plus a text block with the window title, the image size,
and `origin` - the screen coordinate of image pixel (0,0), so any pixel can
be mapped back to a screen position.
max_side downscales large windows to keep image tokens sane (0 disables).
Read-only. Minimised windows and some GPU-rendered surfaces capture blank,
which is reported instead of a useless image.
Inputs: window (required), max_side.
get_ui_tree
Read the UI Automation tree of one window as a compact text list.
`window` is an hwnd from list_windows (e.g. "132456") or a unique part of a
title or process name (e.g. "notepad"). ambiguous matches are an error, not
a guess.
Returns one line per interactive element: path, control type, name,
automation id, screen rect. Invisible and zero-size elements are dropped.
max_depth defaults to 8 and the walk is breadth-first, because shallow
depth limits hide real controls (a calculator loses its keypad at depth 4,
Explorer hides the file list at depth 6). max_nodes caps the walk: dense
windows cost about 2.5 ms per node, so raise it deliberately.
include_text also lists static text elements, which is useful for reading
labels and status messages. include_containers adds List/Tree/Table/DataGrid
nodes: their paths are what inspect_element expects when you want to read a
whole table or list instead of numbering every row.
Read-only. If the window runs elevated while this server does not, Windows
blocks UIA and the tool reports that instead of returning an empty tree.
Inputs: window (required), max_depth, max_nodes, include_text, +1 more. Returns: result
inspect_element
Read the full content of one element, identified by a path.
Call get_ui_tree first, then pass the path of the element you care about
(for example "0.4.0.1"). Use this for elements that hold a lot of content -
a document, a table, a long list, an edit field - where numbering every row
would waste the whole mark budget.
Returns the element's type, name, screen rect, its value or text (up to
max_chars), and a summary of its immediate children. That child summary is
usually enough to understand a table without capturing or marking it.
Paths are positional: if the UI changed between calls, call get_ui_tree
again. Read-only.
Inputs: path (required), window (required), max_chars. Returns: result
list_windows
List visible top-level windows, foreground first.
Returns hwnd, process name, rect in physical screen pixels, whether the
window is covered by another window (occluded), and whether it can only be
read by an elevated server (elevated). Titles are localized by Windows.
Use the hwnd with get_ui_tree and capture_window. hwnds are stable for the
lifetime of a window but change when the app is restarted.
Inputs: include_desktop. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
TDQS
Score is being calculated.
Maintenance
Related MCP Connectors
Eyes and hands on real Windows PCs — observe, click, type via Glasswarp API.
Give your agent eyes: screenshots of any web page it can see, pages as markdown, and PDFs.
Securely control computers you explicitly pair through files, terminals, processes, screenshots, desktop UI/input, clipboard, browser automation, diagnostics, and document tools.
Give agents eyes on any web page: structured context, and changes explained in plain language.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceGives MCP-capable AI tools running on Windows sight of the user's desktop: it captures the window being looked at, any monitor, window or region, and offers OCR, text finding, full-resolution zooming, snips, hotkey marks, visual baseline comparison and waiting for screen changes. Sensitive apps and windows are blacked out before any image reaches the AI client.MIT
- AlicenseNot gradedqualityCmaintenanceEnables MCP clients to read any Windows window as text with coordinates, including control names, roles, clickability, and all on-screen text positions, so even models without vision can understand and act on the desktop. It is read-only, listing windows, finding elements by name, and waiting for text to appear, without clicking, typing, or stealing focus.MIT
- AlicenseNot gradedqualityCmaintenanceLets an AI model see an indexed UI Automation accessibility tree of Windows desktop apps and act on it with precise clicks, typing, key chords, scrolling, dragging, and element value changes, using window captures that work even for occluded windows. Screenshot-based coordinates are rescaled and validated against the target window so input fails loudly instead of hitting the wrong app.MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to read what is on a Windows desktop as text — open windows, their contents, and what changed — using Win32 and UI Automation with optional NPU vision. Secrets, password fields, and private windows are withheld and personal data scrubbed on the user's PC before anything reaches the agent.AGPL 3.0