Single tool 'ask_gemini' has a reasonable name starting with a verb, but the implementation has significant gaps in schema completeness, parameter descriptions, and output documentation. The tool accepts only one parameter ('prompt') with a moderately detailed description (123 characters, within baseline range of 34-392), but lacks comprehensive output schema documentation and error handling guidance. No tool annotations present. The server structure is clean but offers minimal composability or advanced features.

• Add explicit output schema to ask_gemini tool definition. Document return type as object with fields: 'text' (string), 'sources' (string[] from URL context and grounding metadata), 'metadata' (optional object with timestamps, model info). Include pagination guidance if responses can exceed a token limit.
• Enhance 'prompt' parameter description with format constraints: 'String, 1-8000 characters. Include URLs as http:// or https:// prefixed text. URLs will be automatically fetched and analyzed. Separate multiple URLs with newlines.'
• Implement structured error responses. Instead of 'Error: <message>', return JSON with fields: 'error_type' (enum: auth_failed, invalid_input, api_error, timeout, rate_limited), 'user_message' (actionable text for LLM), 'recovery_suggestion' (e.g. 'Check GEMINI_API_KEY environment variable' or 'Try again in 30 seconds').
• Add tool annotations to ask_gemini schema. Set 'readOnlyHint': true (confirmed by Risk: READ_ONLY) and 'idempotentHint': true (query-only, side-effect free). This enables agents to retry confidently.
• Remove advertised 'prompts' capability from Server constructor unless prompts are fully implemented. If prompts are planned, return actual prompt definitions from ListPromptsRequestSchema with descriptions and arguments, and implement GetPromptRequestSchema.
• Validate GEMINI_API_KEY and GEMINI_MODEL inside GeminiClient constructor with detailed error messages. If validation fails, throw an error that includes 'Required environment variables: GEMINI_API_KEY, GEMINI_MODEL (defaults to models/gemini-2.0-flash if unset)'.