State-of-the-art Unified Rust MCP server for Windows 10+ automation and vision.
UltraWin has 8 tools with basic schemas and descriptions, but significant gaps prevent production use. All tools have descriptions (10-80 chars, mostly short), and input schemas are present with proper JSON Schema structure. However, parameter descriptions are almost entirely missing, only 'query' (find_element), 'selector' (web_query), and 'text' (type_text) have descriptions; 'x', 'y', 'button' in click lack any guidance. No output schemas documented. Error handling is minimal, tools return generic 'InternalError' with no recovery guidance. No tool annotations (readOnlyHint/destructiveHint). Naming is verb-first and clear (screenshot, click, type_text), but descriptions are too terse to guide LLM selection (e.g., 'Capture a zero-latency screenshot using DXGI' doesn't explain when to use it vs. read_text). No parameter constraints documented (e.g., button enum is declared but not described; x/y ranges unknown). Composition is reasonable, tools are single-purpose, but no chaining IDs returned in responses (inferred from code structure). Risk classification (READ_ONLY vs WRITE) is present but not exposed in tool definitions.
Click at specific screen coordinates
Find an element by name or ID and return its bounding box
Get properties of the currently focused UI element
Retrieve the accessibility tree (recursive)
Perform hardware-accelerated OCR on the current screen
Capture a zero-latency screenshot using DXGI
Type text into the focused element
Missing parameter descriptions for 5+ parameters (x, y, button in click; no descriptions for any params in screenshot, get_ui_tree, get_focused_element, read_text). LLMs cannot infer parameter meaning from names alone.
No output schemas documented. LLMs cannot plan downstream calls or extract required fields (e.g., what does get_ui_tree return? What fields? What structure?).
Tool descriptions are too terse (10-80 chars, baseline 194 chars). Descriptions do not explain WHEN to use each tool or what it returns. E.g., 'Capture a zero-latency screenshot using DXGI' lacks context for LLM selection vs. read_text.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | <=2025-11-25 | v2 |
Query the browser DOM using a CSS selector (requires --remote-debugging-port=9222)
No error recovery guidance. Tools return generic 'InternalError' with no actionable next steps. E.g., 'Capture failed: ...' tells LLM nothing about retry, fallback, or root cause.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Risk classification exists in metadata but is not exposed to LLM via tool definitions, preventing safe agent planning.