Agnes MCP provides 4 tools with complete input schemas and descriptions, but quality is uneven. Tool names follow verb_noun convention (generate_image, create_video_task, get_video_result, generate_video). Descriptions are present but brief (50-100 chars), lacking context on when to use each tool or dependencies between them. Schemas are well-formed with typed parameters and descriptions, but output schemas are not documented. Error handling is minimal, no recovery guidance or actionable error messages. The server lacks tool annotations (readOnlyHint, destructiveHint) despite having clear read/write semantics. No pagination support despite potential for large result sets.

• Document output schemas for all tools. Example: generate_image returns {url: string, local_path: string, model: string, size: string, n: int, images: [{url, local_path}]}. This enables LLMs to chain calls (e.g., pass returned image URLs to create_video_task).
• Add tool annotations: mark generate_image, create_video_task, generate_video with destructiveHint=true; mark get_video_result with readOnlyHint=true. This guides LLM retry logic and side-effect reasoning.
• Enhance tool descriptions with selection context. Example: 'Generate images from text. Use this for text-to-image or image-to-image (via images param). For video generation, use create_video_task or generate_video instead.'
• Add enum constraints to schema for return_mode ('url'|'b64') and mode (if applicable). Replace free-form strings with formal constraints to prevent hallucinated values.
• Implement actionable error messages. Instead of 'API error 400: ...', return 'Invalid prompt: too long (max 2000 chars)' or 'Rate limited. Retry in 60 seconds.' This enables LLM self-correction.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry