The server provides 6 tools for MongoDB data access and NWB file handling. Most tools have descriptions and basic schemas, but quality is inconsistent. Tool naming is clear (verb-first: get_, identify_), but parameter descriptions vary significantly in completeness. No output schemas are documented. Error handling returns raw exception messages rather than actionable guidance. The server reads from external databases and file systems but lacks input validation, type constraints, and recovery instructions.

• Document output schemas for all tools. Example for get_records: 'Returns: list of dicts, each with fields {_id: str, subject: {sex: str, ...}, ...}. Includes all fields unless projection specified.'
• Replace generic exception messages with actionable errors. Example: 'Invalid filter syntax. Expected MongoDB query dict, e.g. {"subject.sex": "Male"}. If error persists, check schema with describe_schema() or consult MongoDB documentation.'
• Add formal constraints to dict parameters. Consider accepting structured input or validating common MongoDB patterns server-side. For aggregation_retrieval, document the expected structure with a concrete example: [{"$match": {...}}, {"$project": {...}}].
• Expand get_summary description to 50 - 150 chars. Include: 'Returns an AI-generated natural-language summary of a data asset (metadata, subject info, methods). Use when you need a quick overview instead of raw documents. Requires _id of an existing asset.'
• Introduce pagination. Add optional 'offset' (default 0) and 'limit' (default 20, max 100) to get_records and aggregation_retrieval. Return {items: [...], total: int, has_more: bool}.
• Add input validation and return clear error messages. Example: 'Invalid _id format. _id must be a non-empty string; got empty string.'
• Document the structure returned by identify_nwb_contents_in_code_ocean and identify_nwb_contents_with_s3_link. What fields are in the NWBZarrIO object? What can the LLM query next?
• Consider splitting tools that conflate concerns. identify_nwb_contents_in_code_ocean both searches directories AND loads the file, consider separate 'search_nwb_files' and 'load_nwb_file' tools.