Three tools with clear action verbs and basic parameter descriptions. All tools have documented input schemas via Zod with enums for constrained inputs (models, aspect ratios, styles). However, parameter descriptions are minimal (under 50 chars each), output schemas are undocumented in code (only inferred from JSON.stringify calls), and error handling is absent. Tool descriptions are adequate (120-180 chars) but lack WHEN to use hints or prerequisites. No documentation of return field structure for downstream tool composition. generate_image modifies state but description does not explicitly say so.

• Document output schemas explicitly. Add JSDoc or inline comments for each tool showing the structure of returned JSON fields (id, status, credits_charged, etc.) so LLMs know what to extract.
• Expand parameter descriptions to 50-100 chars with format and constraint details. E.g., 'prompt: A detailed image description (1-1000 chars). More specific prompts yield better results. Avoid offensive content.' instead of just 'Image description / prompt'.
• Update generate_image description to state: 'Generates an AI image and saves it to disk. MODIFIES STATE: charges credits and creates file. Provide client with confirmation step before calling.' This flags the tool as a command-tool.
• Add error handling with recovery hints. Catch missing API key, insufficient balance, network errors, and return structured errors like '{"error": "insufficient_credits", "message": "Check balance with check_balance tool first.", "retryable": false}'.
• Add tool annotations via server.tool() metadata: mark list_models and check_balance with readOnlyHint: true, and generate_image with destructiveHint: true to enable client-side filtering.
• Document idempotency guarantees. Is calling generate_image twice with identical params deterministic or will it create two images? State this explicitly.
• Add 'aspect_ratio' and 'style' descriptions that explain their visual impact. E.g., 'aspect_ratio: Canvas dimensions. 1:1 = square, 16:9 = wide landscape, 9:16 = tall portrait. Default 1:1.'
• Return a structured response object with documented fields. Instead of JSON.stringify, return {content: [{type: 'text', text: JSON.stringify({...})}]} with clear field meanings in tool description.