MCP server for fetching trending posts from Xiaohongshu (Little Red Book), a Chinese social media platform
This MCP server exposes a single tool for fetching Xiaohongshu (XHS) posts. While the tool is functionally implemented with proper retry logic and browser automation, the definition quality is significantly hampered by incomplete schema documentation, sparse parameter validation guidance, and missing output schema specification. The tool name is reasonably descriptive ('fetch_xhs_hot_post'), but critical schema and description deficiencies prevent LLMs from understanding the full interface. The server returns unstructured string output rather than typed objects, limiting downstream tool composition. Based on rubric baselines (avg tool description 194 chars, 100% of A+ tools have documented return types), this server falls well short of production quality.
从 小红书 获取爆款帖子数据
Return type is unstructured string (union of str | Exception) with no output schema documentation. LLMs cannot parse or chain results downstream. The tool returns markdown text, but no schema specifies fields, structure, or pagination.
Parameter 'search' lacks format constraints in schema. Description says '长度不超过15个字' (max 15 chars), but this validation rule is in natural language only, not expressed as a JSON Schema constraint (maxLength). LLMs cannot parse natural-language constraints reliably.
Parameter 'limit' has a constraint ('不超过10', max 10) stated only in natural language description, not in the JSON Schema (maximum: 10). Schema shows only default:5, type:integer with no bounds.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | F | 40 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 41 | - | v1 |
Tool description is brief (67 chars: '从 小红书 获取爆款帖子数据') and lacks contextual guidance. Does not explain WHEN to call this tool vs alternatives, what data structure is returned, expected latency, or error conditions. Baseline for A+ tools is 194 chars avg with clear WHAT/WHEN/WHY.
Error handling returns bare Exception objects or generic strings ('无数据', 'Error: 搜索字符串长度不能超过15个字'). No structured error responses. LLMs receive unstructured error text with no guidance on recovery actions (retry, fallback, user action needed).
Tool uses browser automation (Drission+Chromium) which may require manual QR-code login (seen in logic.py:new_browser_and_search sending Dingtalk notifications for QR codes). This introduces statefulness and blocking I/O that conflicts with stateless MCP request semantics. No documentation of login prerequisites or session handling.
No tool annotations present (readOnlyHint, destructiveHint, idempotentHint). This tool is READ_ONLY per metadata but the MCP tool definition does not declare this. Current spec encourages explicit annotations for LLM planning.