An MCP (Model Context Protocol) server for asynchronous image generation using the Tu-zi API
Server has 4 tools with explicit schemas and descriptions. Naming follows verb_noun pattern (submit_*, wait_*). Descriptions are present (100-200 chars) and explain WHEN to use tools. However, parameter descriptions lack actionable constraints (e.g., no guidance on aspect ratio format, no validation rules for paths). Output schemas are not documented. Error handling is minimal, ToolError is raised but does not guide recovery. No tool annotations (readOnlyHint/destructiveHint). Reference image path parsing is undocumented (comma-separated string is non-standard). Overall: solid foundation but missing production-grade polish.
Submit a Gemini image generation task. Use wait_tasks() to wait for all submitted tasks to complete.
Submit an async GPT image generation task. Use wait_tasks() to wait for all submitted tasks to complete.
Submit a Seedream (即梦) image generation/editing task. Suitable for Chinese-context tasks. For generating multiple images while keeping same style (n>1), use the prompt format: '[General style description] Create separate images: Image 1: [description], Image 2: [description], Image 3: [description], Image 4: [description]' Output files are saved with _1, _2, _3, _4 suffixes when n>1. Use wait_tasks() to wait for completion.
Wait for all previously submitted image generation tasks to complete.
Output schemas not documented. Tools return ToolResult with TextContent but no structured response schema is defined. LLMs cannot plan downstream calls or extract task IDs reliably.
Parameter descriptions lack actionable constraints. 'prompt' parameter says 'Must include aspect ratio' but does not specify format (e.g., '1:1', '3:2') or consequences of omission. 'output_path' says 'Absolute path' but does not validate or document error handling for invalid paths.
Error handling is minimal. ToolError is raised with generic messages ('Failed to submit task') but does not guide recovery. No distinction between retryable errors (network timeout) and user-fixable errors (invalid path, missing aspect ratio).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | 2026-07-28+ | v2 |
Reference image path parsing is non-standard. Comma-separated string ('path1,path2') is undocumented and error-prone. No validation that paths exist or are readable. LLMs may pass invalid paths without feedback.
Tool annotations missing. No readOnlyHint, destructiveHint, or idempotentHint declared. Agents cannot determine which tools are safe to retry or have side effects.