The GTEx server has 18 tool definitions with reasonable naming conventions (verb-noun structure) and explicit input schemas, but suffers from significant quality gaps. All tools have descriptions and schemas, but many descriptions are too brief (under 50 chars) to fully guide LLM selection. Parameter descriptions vary widely in quality, some are specific (e.g., 'GENCODE gene ID (e.g., ENSG00000223972.5)'), others are generic ('Array of GENCODE gene IDs'). Tools 16 and 17 are near-duplicates of tools 8 and 9 (both named 'get_eqtl_genes' and 'get_single_tissue_eqtls'), indicating composition issues. Output schemas are completely undocumented, no tool specifies what fields the response contains, forcing LLMs to discover structure at runtime. Error handling is absent: no tool description indicates what errors might occur or how to recover. Most critically, several tools have duplicate names (tools 8 & 16, tools 9 & 17, tools 11 & 18), violating basic tool composition principles.

• CRITICAL: Rename duplicate tools. Tools 8 & 16 conflict, rename one to 'get_eqtl_genes_by_region' (region query) and keep the other as 'get_eqtl_genes_by_tissue' (tissue-specific). Same for tools 9 & 17 and tools 11 & 18. Names must unambiguously distinguish intent.
• CRITICAL: Document output schemas for all 18 tools. Add a 'returns' section to each tool definition specifying response structure. Example: 'Returns: { tissues: [{name: string, medianExpression: number}], geneId: string, datasetId: string }'. This is essential for LLM composition.
• Expand tool descriptions to 100 - 200 chars with clear guidance on when to call each. Example: 'Get gene expression data across all tissues for a single gene. Use this to explore a gene's expression landscape; use get_median_gene_expression if you only need tissue-level summaries.' This disambiguates from related tools.
• Standardize pagination: all tools returning lists should support 'page' and 'itemsPerPage' parameters with consistent naming. Add a 'total' or 'nextPage' field to responses so LLMs know when to fetch more.
• Add error guidance to each description. Example: 'If gencodeId is invalid, returns 400 with message 'Gene ID not found. Search with search_genes() to find valid IDs.' This teaches LLMs to self-correct.
• Document array cardinality in parameters. For 'gencodeIds' (tools 5, 6), specify min/max length (e.g., 'Array of 1 - 100 GENCODE gene IDs'). LLMs will pass absurd values without bounds.
• Tools 16 & 17: add at least one required parameter (e.g., require at least one of 'tissueSiteDetailId', 'geneIds', 'variantIds'). A tool with no required params is ambiguous and invites misuse.