Server exposes 6 core tools + 2 semantic tools with minimal schema documentation. Core tools have brief descriptions (10-60 chars) but lack parameter descriptions entirely. Input schemas are present but sparse: set_widget and click accept only identifier/value with no type constraints, enums, or validation guidance. No output schemas documented. Error handling is absent, tools return raw engine results without recovery guidance. Naming is verb-first (list_, get_, set_, click, read_, reset_) which is good, but descriptions are too terse to guide LLM selection. No tool annotations (readOnlyHint/destructiveHint) despite clear read/write risk classification in metadata.

• Add parameter descriptions to set_widget and click: 'identifier (string): Widget key or label to target. Use list_widgets() to discover available keys.' and 'value (any): New value to set. Type depends on widget (e.g. string for text input, bool for checkbox, int for slider).'
• Document output schemas for all tools. E.g. list_widgets returns {widgets: [{key: string, label: string, type: string, schema: object}], total: int}. Include this in docstrings or a separate schema registry.
• Expand tool descriptions to 100-150 chars. E.g. 'List all interactive widgets in the app with their MCP tool schemas. Call this first to discover available controls before calling set_widget or click.'
• Add error handling wrapper: catch engine exceptions and return {error: string, suggestion: string, retryable: bool}. E.g. 'Widget key "invalid_key" not found. Available keys: [list]. Try set_widget with a valid key.'
• Implement tool annotations via FastMCP: mark read-only tools (list_widgets, get_layout, read_output, get_state) with readOnlyHint=true and destructive tools (set_widget, click, reset_*) with destructiveHint=true to guide agent planning.