CHGIS时空地名查询API的MCP服务器包装器 (MCP server wrapper for CHGIS spatio-temporal place name query API)
The CHGIS server has well-structured tool definitions with proper JSON schemas, clear naming conventions using action verbs, and reasonable descriptions. However, there are significant gaps in error guidance, output schema documentation, and parameter validation messaging that prevent this from reaching a higher tier. All three tools follow verb_noun naming (search_*, get_*), have input schemas with proper type constraints, and descriptions between 40-80 characters, reasonable but minimal. The critical issue is that output schemas are not documented anywhere in the code, error messages lack recovery guidance, and parameter descriptions are sparse on constraints (e.g., what 'feature_type' values are actually valid beyond the example 'zhou, xian, fu'). The helper functions formatPlaceDetails() and formatSearchResults() exist but their output structure is not documented for LLMs to plan downstream calls.
Get historical context and hierarchical relationships of a place
Query historical place details by unique ID
Search places by name, year, administrative level, etc.
Output schemas are not documented. Code contains formatPlaceDetails(), formatSearchResults(), and formatHistoricalContext() but their return structures are never shown to the LLM. This prevents downstream tool composition and forces LLMs to guess what fields are available. Agents cannot plan multi-step calls without knowing what data each tool returns.
Error messages lack recovery guidance. When validation fails (e.g., 'At least one search parameter must be provided'), the code throws McpError but does not suggest which parameters are valid or what to try next. Per the rubric, errors must answer: 'what went wrong?' and 'what should the LLM do next?' Currently only the first question is answered.
Parameter descriptions are incomplete on constraints. 'feature_type' description says 'Administrative level type (e.g., zhou, xian, fu)' but does not declare this as an enum or state the full list of valid values. 'source' is properly constrained with enum=['CHGIS','RAS'], but feature_type is left as free-form string. This invites hallucinated values from the LLM.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 0 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 45 | - | v1 |
Parameter descriptions lack actionable format guidance. 'name' in search_places says 'Place name (supports Chinese, Pinyin, etc.)' but does not state: Do partial matches work? Are diacritics required? How long can names be? 'parent' says 'Parent place or administrative division', what format is expected? A place ID? A name? This ambiguity forces LLMs to guess.
Tool description for search_places lacks clear distinction from search_place_by_id. Both search places; one by ID, one by filters. The descriptions do not clearly state: 'Use search_place_by_id when you have a specific hvd_* ID. Use search_places when you need to discover places by name, year, or type.' This risks LLM confusion between the two tools.
No pagination or result limits declared. The searchPlaces() call passes arbitrary parameters to an external API but the tool description does not state: 'Returns up to N results' or 'Use limit/offset for pagination.' If the CHGIS API can return hundreds of places, unpaginated results blow the context window. The code does not enforce a cap.