Server provides 6 tools with complete input schemas and descriptions. Tool names follow verb_noun convention (gemini_generate, gemini_chat, gemini_vision, gemini_embed, gemini_count_tokens, gemini_list_models). Descriptions are present and contextual (avg ~120 chars), exceeding the 10-char minimum but below the 50-200 char LLM-optimized range. All parameters have type definitions and descriptions. However, output schemas are not documented, responses are described only in prose. Error handling is present but generic (maps HTTP status codes without recovery guidance). No pagination support despite tools that could return large results (gemini_list_models, gemini_embed). Pydantic models enforce input validation, but error messages returned to LLM are not actionable enough to guide self-correction.

• Document output schemas for all tools. E.g., gemini_generate returns {text: string, usage: {input_tokens: int, output_tokens: int}}. gemini_embed returns {embeddings: [{values: float[]}]}. This enables LLMs to plan chained calls.
• Enhance error messages with recovery guidance. Replace 'Error: HTTP 400: Bad request' with 'Invalid prompt format. Ensure prompt is a non-empty string under 100,000 characters. If using JSON mode, wrap expected JSON structure in triple backticks.'
• Add pagination to gemini_embed and gemini_list_models. Accept limit (default 20, max 100) and offset/cursor parameters. Return total_count and next_cursor in response.
• Clarify tool selection in descriptions. E.g., gemini_generate: 'Use for single-turn text generation when you do not need conversation history.' gemini_chat: 'Use for multi-turn conversations where you need to maintain message history across turns.'
• Add per-parameter validation error messages. When temperature is out of range, return 'Invalid temperature: got 2.5, must be between 0.0 and 2.0' instead of generic validation error.