11 tools with basic schemas and descriptions, but significant gaps. All tools have descriptions (10-50 chars, below the 50-200 char baseline for LLM optimization). Input schemas present for all tools with proper type definitions, but output schemas are completely undocumented, LLMs cannot plan downstream calls or extract structured data. No error handling guidance, no parameter constraints (enums), no pagination support despite list-returning tools. Tool naming is verb-first and clear (get_*, search_*, accept_*, etc.), but descriptions are too terse to guide LLM selection. No tool annotations (readOnlyHint/destructiveHint). Security: no audit logging visible, no permission gates documented.

• Document output schemas for all tools. E.g., get_conversations should return {conversations: [{id: string, name: string, last_message_time: number, message_count: number}], total_count: number}. This enables LLMs to chain tools and extract IDs for follow-up calls.
• Expand descriptions to 50-200 chars with WHAT/WHEN/WHY. Example: 'Get all conversations from Session database. Use this to discover available chats before fetching messages. Returns conversation IDs, names, and metadata needed for get_messages() calls.'
• Add pagination to list tools: get_conversations(limit=20, offset=0), get_messages(conversation_id, limit=50, offset=0), search_messages(query, limit=20, offset=0). Return total_count and next_offset in responses.
• Add enum constraints to format parameter in export_conversation: {type: 'string', enum: ['json', 'csv', 'html']}. Add descriptions to include_attachments explaining the impact on file size.
• Add tool annotations via fastmcp: mark get_* tools with readOnlyHint=true; mark accept_request, decline_request, block_request with destructiveHint=true and idempotentHint=false (since accepting twice may fail). This guides LLM retry logic.