Server has 5 tools with complete descriptions and basic schemas. Naming follows verb_noun convention (debate_with_gemini, reset_debate, get_transcript, judge_debate, list_debates). Descriptions are detailed (200-400 chars) and explain WHEN/WHY to use each tool. However, parameter descriptions lack format constraints (e.g., conversation_id has no length/pattern spec), error handling is minimal (returns string errors without recovery guidance), and output schemas are undocumented. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Composition is sound, tools chain logically (debate → get_transcript → judge_debate). Missing pagination for list_debates and no structured output format.

• Add output schema documentation for each tool. E.g., debate_with_gemini returns {text: string, error?: string}; judge_debate returns {winner: 'claude'|'gemini', claude_score: number, gemini_score: number, reasoning: string}.
• Add parameter constraints: conversation_id (1-64 alphanumeric chars, pattern: ^[a-zA-Z0-9_-]+$), persona (max 500 chars), message (min 1, max 5000 chars).
• Implement error classification. Return structured errors: {error: string, code: 'rate_limit'|'auth_failed'|'invalid_input'|'server_error', retryable: boolean, suggestion: string}.
• Add tool annotations via FastMCP. Mark reset_debate with @mcp.tool(annotations={'destructive': True}); mark read-only tools with readOnlyHint.
• Add pagination to list_debates: accept offset (default 0) and limit (default 20, max 100) parameters; return {conversation_ids: [string], total: number, offset: number, limit: number}.