Two tools with well-structured schemas and clear error handling, but descriptions lack depth and parameter documentation is incomplete. desktop_observe has a complex oneOf schema with good type coverage; desktop_act similarly uses oneOf for action variants. Both tools define error cases with actionable messages (STALE_SNAPSHOT, UNSUPPORTED_KEY, etc.). However, descriptions are terse (under 100 chars), parameter descriptions are absent or minimal, and output schemas lack explicit documentation. The server follows good practices for input validation and error classification but falls short of LLM-optimized descriptions that explain WHEN and WHY to use each tool.

• Expand tool descriptions to 100-200 characters. For desktop_observe: 'Capture the current desktop, active window, or specific output region as a JPEG. Returns snapshot_id (for use in desktop_act), image dimensions, and encoded image data. Call this first to see the current state, then use snapshot_id in subsequent actions.' For desktop_act: 'Execute a single input action (click, move, type, or key press) on the desktop. Requires a valid snapshot_id from a recent desktop_observe call. Actions are atomic, one call = one action. Returns success or error with recovery guidance.'
• Add per-parameter descriptions in the schema. For 'target' in desktop_observe, document each variant: 'kind=desktop: capture full screen. kind=active_window: capture focused window only. kind=output: capture specific output region (requires output parameter with region identifier).' For 'action' in desktop_act, document each action type and required fields.
• Document the output schema explicitly. Add a docstring or schema comment showing: desktop_observe returns {snapshot_id: string, encoded_size: {width: int, height: int}, image: base64-encoded JPEG}. desktop_act returns {success: bool, error?: string}. Include this in tool descriptions or a separate schema reference.
• Add tool annotations to the schema. Mark desktop_observe with readOnlyHint=true. Mark desktop_act with destructiveHint=true (since it modifies desktop state) and idempotentHint=false (since repeated clicks have different effects). This helps agents reason about retry safety.