Server has 9 tools with explicit schemas and descriptions. Naming follows verb_noun convention (list_models, get_credits, generate, etc.). Descriptions are present and actionable (avg ~120 chars), explaining WHAT and WHEN. Input schemas use Zod with proper types and enums. Tool annotations (readOnlyHint, destructiveHint) are declared. However, several parameters lack descriptions (e.g., 'limit' in list_generations has a description but others are sparse), and output schemas are not documented, LLMs cannot predict response structure. Error handling returns structured objects but lacks recovery guidance. Security: API keys injected server-side (good), but no explicit permission gates or audit trail declarations.

• Document output schemas for all tools. E.g., list_models should declare: {models: [{id: string, name: string, creditCost: number, workflows: string[]}], total: number}. This enables LLMs to plan downstream calls.
• Enhance list_generations description: 'List all generations for this API key, with pagination. Returns generation IDs, status, model, and timestamps. Use limit (1-100, default 20) and offset for pagination. Call this to discover available generations before calling get_generation for details.'
• Add recovery guidance to error responses. E.g., if generate fails with 'Insufficient credits', return: {error: 'Insufficient credits (have 5, need 10)', suggestion: 'Call get_credits() to check balance, or upgrade your plan'}.
• Document mutual exclusivity in upload_input: 'Either filePath (local file relative to upload root) OR imageBase64 (base64-encoded image data) must be provided, not both. If both are present, imageBase64 takes precedence.'
• Add idempotency guidance: declare which tools are safe to retry (all read-only tools, generate with wait:true) vs which have side effects (cancel_generation, upload_input). This helps agents recover from transient failures.