MCP agent that translates natural language questions into accurate, explainable, and reproducible data insights
This server has foundational tool definitions but significant gaps prevent production use. All 9 tools have descriptions and basic naming conventions, but 7 of 9 tools lack proper input schemas in the source code. The visible code shows function signatures with parameters, but no explicit JSON Schema definitions. Parameter descriptions exist in docstrings but are not formally registered with type information, making LLM invocation unreliable. Error handling is present but generic (returns strings rather than structured errors). The server uses in-memory stores and mock data, suitable for demonstration but not production deployment. Naming follows verb_object conventions (ask_question, generate_query, validate_query) which is positive, but several tools have overlapping responsibilities that could be split (ask_question combines parsing, query generation, and result formatting).
Ask a data question in natural language and get an explainable answer.
Assess data quality for a specific metric.
Compare two metrics side-by-side with correlation analysis.
Explain a query result with context, comparisons, and interpretation.
Generate a SQL query for a specific metric with structured parameters.
List all saved query templates.
Save a SQL query as a reusable template for future use.
No formal JSON Schema definitions visible in source code. All 9 tools show function signatures with type hints in Python, but no explicit schema registration via @mcp.tool decorator parameters. JSON Schema is required for LLM tool invocation; without it, clients cannot validate inputs before calling the tool.
Output schemas not documented. Tools return plain strings or unstructured text (e.g., ask_question returns a multi-line formatted string). LLMs cannot parse these results reliably to extract structured data for downstream tool chains. No pagination, field documentation, or typed response schemas visible.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 44 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 41 | 1.1.2+ | v1 |
Generate follow-up questions based on an initial question.
Validate a SQL query for syntax, performance, and data quality issues.
Error handling returns free-form strings instead of structured error responses. Example: validate_query and generate_query return plain text error messages. LLMs cannot categorize errors as retryable, user-fixable, or fatal. No recovery guidance or actionable next steps provided.
ask_question combines multiple responsibilities: parsing questions, generating SQL, executing queries, assessing quality, and generating follow-ups. This violates single-responsibility pattern. Agents cannot compose these steps independently or retry individual parts. Should be split into separate tools.
Parameter descriptions in docstrings use example values ('e.g., last 7 days', 'e.g., active_users'). LLMs tend to reuse example values literally, causing failures with real data. Should use enums or formal constraints instead.
metric, time_period, and filters parameters accept free-form strings. No enum constraints for metric names. Invalid metric inputs return a suggestion, but the error does not explicitly state which metrics are valid or provide programmatic choices for LLM recovery.
In-memory data stores (QUERY_HISTORY, QUERY_TEMPLATES, METRIC_DEFINITIONS) are non-persistent and simulate query execution. Production deployment requires integration with real databases, but no configuration, environment variables, or connection abstraction is visible. Mock data masks real-world integration gaps.
No input validation or sanitization. SQL queries and filter parameters are concatenated directly into WHERE clauses without escaping. LLMs can be tricked into injecting SQL via 'filters' or 'group_by' parameters. Vulnerable to prompt injection attacks.
No pagination or result limits. list_templates returns all templates without pagination. Large result sets could exhaust context windows. Tools like ask_question and suggest_followups return unbounded lists of follow-up suggestions.
Tool annotations (readOnlyHint, destructiveHint, idempotentHint) are not declared. Most tools are read-only, but no formal annotations guide LLM behavior. save_query_template is destructive (writes to template store) but has no explicit destructiveHint or confirmation mechanism.