Dropscan MCP has 13 well-named, read-only tools with consistent verb_noun naming (list_*, get_*). All tools have descriptions (avg ~80 chars, within baseline 34-392). Input schemas are present and typed for all tools. However, descriptions are terse and lack context for LLM selection, they state WHAT (e.g., 'List all scanboxes') but not WHEN or WHY to use each tool. Parameter descriptions are minimal (e.g., 'Scanbox id' vs. 'The unique identifier for the scanbox; required to filter mailings'). Output schemas are undocumented, LLMs cannot predict response structure. Error handling returns generic dicts without recovery guidance. No tool annotations (readOnlyHint, idempotentHint) despite all being read-only. Security is strong (token via env var, no secrets in params), but composition lacks pagination guidance for list tools.

• Add output schema documentation to every tool. For list_mailings, document: 'Returns {mailings: [{uuid, sender, subject, status, received_at, ...}], total_count, next_cursor}'. For get_mailing, document nested fields (extracted_text, pages, etc.).
• Expand tool descriptions to 100-200 chars. E.g., 'List all scanboxes on your Dropscan account. Call this first to discover available mailboxes before listing mailings. Returns scanbox IDs, names, and status.'
• Replace example values with format constraints. Change 'Scanbox id (e.g. "12345")' to 'Scanbox ID (numeric string, 1-10 digits)'. Add enum for current_status: ['received', 'scan_requested', 'scanned', 'destroy_requested', 'destroyed', 'forward_requested', 'forwarded'].
• Add tool annotations via FastMCP. Mark all tools with readOnlyHint=true and idempotentHint=true so LLMs know they are safe to retry. E.g., @mcp.tool(readOnlyHint=True, idempotentHint=True).
• Enhance error responses with recovery hints. E.g., if scanbox_id not found, return {'error': 'Scanbox not found', 'hint': 'Call list_scanboxes() to see available scanboxes', 'available_ids': [...]}