anndata-mcp provides 4 well-named, READ_ONLY tools with complete input schemas and descriptions. Tool names follow verb_noun convention (get_*, view_*, locate_*), which is excellent for LLM parsing. All tools have clear descriptions (122-190 chars), meeting the baseline of 194 chars average. Input schemas are detailed with proper type declarations, enums, and nested parameter support. However, output schemas are not documented, LLMs cannot determine what fields to expect from results. Error handling guidance is absent; tools lack recovery instructions or categorization. Parameter descriptions could be more prescriptive about expected formats and edge cases. No tool annotations (readOnlyHint, etc.) despite all being read-only operations.

• Document the output schema for each tool. For get_descriptive_stats, specify the structure of the returned statistics object (e.g., {mean: float, std: float, count: int, value_counts: {category: count, ...}}). For view_raw_data, document the sliced array/dataframe format.
• Add error handling guidance to each tool description. Example: 'If the file path is invalid, use locate_anndata_stores() to discover available files. If the key does not exist, the tool will return an error listing available keys.'
• Provide concrete examples for complex parameters. For get_descriptive_stats, add: 'key=null for simple attributes, key=["level1", "level2"] for nested access, columns_or_genes=["CD4", "CD8*"] to select specific genes or use glob patterns.'
• Explicitly document parameter dependencies. For filter parameters, state: 'To filter, provide all four parameters together: filter_attribute (obs or var), filter_column (column name), filter_operator (==, !=, etc.), and filter_value (literal or array for isin/notin).'
• Add readOnlyHint=true to all tool definitions in the schema to signal these are safe, idempotent operations that agents can retry without side effects.
• Document output limits in tool descriptions. Specify max rows returned, truncation behavior for large results, and any default limits (e.g., max 1000 genes in descriptive stats).
• For locate_anndata_stores, document the return format: is it a list of paths, a structured object with path/size/date metadata, or something else? Agents need to know how to extract file paths from the result.