Server has 10 tools with complete schemas and descriptions, but quality is uneven. Tool names follow verb_noun convention well (get_*, add_*, start_*, save_*). Descriptions are present but often generic or lack actionable context. Parameters have types and descriptions, but many lack constraints (enums, ranges). Error handling returns JSON with error codes but lacks recovery guidance. No tool annotations (readOnlyHint/destructiveHint). Output schemas are not formally documented. The server accepts opaque IDs (host_id format 'https:example.com:443') without accepting human-friendly alternatives, forcing extra lookup calls.

• Add tool annotations: @mcp.tool(readOnlyHint=True) for get_* tools; @mcp.tool(destructiveHint=True) for add_host, save_token, start_auth. This signals to LLMs which tools are safe to retry.
• Convert free-form string parameters to enums. Define order_by as enum ['TOTAL_SHOWS', 'TOTAL_CLICKS']; device_type_indicator as enum ['ALL', 'DESKTOP', 'MOBILE', 'PHONE', 'TABLET']; query_indicator as enum or constrained string with pattern.
• Add a resolve_host_id(site_name_or_url: str) -> str tool that accepts human-friendly site names/URLs and returns the opaque host_id. This eliminates the need for agents to call get_hosts as a discovery step.
• Enhance error messages with recovery guidance. Instead of 'AUTH_ERROR', return 'Authentication failed. Call start_auth(client_id=...) to re-authenticate, then retry.' This guides the agent's next action.
• Document output schemas in tool descriptions or as a separate schema field. E.g. 'Returns: {user_id: string, hosts: [{host_id: string, host_url: string, sqi: number}]}' so LLMs know what fields to expect.