MongoDB MCP server has fundamental definition gaps typical of early-stage implementations. While tool names follow verb-noun conventions (list_*, find_*, insert_*, count_*, aggregate, distinct), critical parameter descriptions are either missing or minimal. Most tools lack documented output schemas, and error handling guidance is absent. The descriptions themselves are non-empty but many fall below the 50-character baseline needed for effective LLM selection reasoning. Parameter descriptions exist for some inputs but lack constraint information (ranges, formats, allowed values). None of the 8 tools include output schema documentation, making downstream tool composition risky.

• Document the complete output schema for every tool. Example for find_documents: 'Returns a list of serialized MongoDB documents. Each document is a dict with ObjectId fields converted to ISO 8601 strings (not raw ObjectId objects). Includes all fields from the collection; use filter to limit which documents are returned.'
• Add constraint documentation to parameters. Example: 'limit (1-10000, default 10), maximum documents to return' and 'filter (MongoDB query syntax, e.g. {"status": "active"}, validated server-side), selection criteria'
• For find_documents and aggregate, add a 'total_count' field to responses so the agent knows whether more results exist. Return {'documents': [...], 'count': 42, 'has_more': true} instead of bare list.
• Add error recovery guidance to tool descriptions. Example: 'If you receive an authentication error, verify MONGO_URI is set correctly. If a database/collection does not exist, call list_databases() and list_collections() first.'
• For insert_document, return the full inserted document (not just ID) so agents can verify success without a follow-up call. Also add a confirmation step or document that this is irreversible.
• Define what happens when find_documents filter is malformed (e.g. {"$where": "malicious code"}). Document validation rules and whether injection attacks are caught server-side.
• Add a parameter 'filter' to find_one explicitly marked as optional with default '{}' so agent knows empty filter is valid (returns first document).
• For aggregate, document the expected pipeline syntax and mention MongoDB aggregation framework versioning, which varies by server version. Note that some $stages may not be available on all versions.