Tiny MCP server for local Windows desktop control
8 tools with basic schemas and descriptions, but significant gaps. All tools have descriptions (10-50 chars, mostly short) and input schemas with types. However, descriptions lack context on WHEN to use tools, prerequisites, or state-modification warnings. Parameters are minimally documented, many lack descriptions entirely (e.g., 'x', 'y' in move_mouse, click, scroll). No output schemas documented. Error handling absent. Tools follow verb_noun naming (screenshot, cursor_position, move_mouse, click, scroll, press_key, hotkey, type_text) but descriptions are generic and do not guide LLM selection. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles (READ_ONLY vs WRITE). Composition is sound, each tool does one thing, but descriptions do not explain dependencies or chaining.
Click the mouse at a position
Get the current cursor position
Press a hotkey combination
Move the cursor to a specific position
Press a single key
Capture a screenshot of the current screen
Scroll the mouse wheel
Type text into the focused application
No output schemas documented. LLMs cannot infer what fields to expect from tool responses (e.g., screenshot returns base64 image data, cursor_position returns {x, y}). This forces agents to guess downstream field names and risks failed chaining.
Parameter descriptions missing or trivial. 'x' and 'y' in move_mouse, click, scroll lack descriptions. 'key' in press_key has no description. 'keys' in hotkey is vague ('Keys to press as a combination', does not explain format or valid values). LLMs cannot infer parameter semantics from names alone.
Tool descriptions do not indicate state modification. Tools like click, scroll, press_key, hotkey, type_text are WRITE operations but descriptions do not warn agents of side effects or irreversibility. Agents cannot determine if a call is safe to retry.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 42 | 2026-07-28+ | v2 |
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). Risk field is present in metadata but not exposed in tool schema. LLMs cannot see which tools are safe (screenshot, cursor_position) vs destructive (click, type_text).
No error handling or recovery guidance. Tools do not document failure modes (e.g., what if screenshot fails? what if click misses the target?). Agents have no guidance on retrying or alternative actions.