MCP server for querying inmydata business intelligence and data warehouse, with support for structured queries, top-N rankings, free-form natural language questions, and agentic RAG knowledge base queries
The server defines 7 tools with action verbs and descriptions, but quality is uneven. Most tool descriptions are verbose and function more as implementation notes than LLM guidance. Parameter schemas are present but lack depth, many parameters have minimal descriptions. Output schemas are partially documented in description text rather than formal schema definitions. Error handling and recovery guidance are weak. The agentic_rag_tool shows good structure and error handling in the code, but the overall design prioritizes implementation details over agent-friendly composition.
Query the tenant's Agentic RAG knowledge base with a natural-language question. Returns an answer grounded in the ingested documents plus the source chunks used. Use this when the user asks about information that would live in the tenant's knowledge base. Do not use for live data, general world knowledge, or anything outside the ingested corpus.
SLOW / EXPENSIVE (fallback). Use ONLY when a request cannot be expressed with get_rows or get_top_n. If the request names explicit fields, filters, years, or dimensions, prefer the fast tools above.
Get all financial periods available in the data warehouse
FAST PATH (recommended). Use when the request names specific fields and simple filters (no free-form reasoning). Returns rows immediately from the warehouse; far faster and cheaper than get_answer. If the output json contains property row_count wil a value less than or equal to 10 then the data property can be used as the answer. If the output json contains a non blank value for the instance_id property then the data property will only contain a sample of the data and the full data set can be found in a table named my_table in a DuckDB database file saved on disk. In that case you MUST use the query_results_fast tool to query the results with SQL to get the data you need to answer the question. This is the only way to access larger datasets.
Output schemas are partially or entirely undocumented. Tools like get_rows_fast, get_top_n_fast, query_results_fast, and get_answer_slow reference output fields (row_count, instance_id, data) in description text but lack formal JSON Schema definitions for response shapes. This forces LLMs to infer structure and plan poorly for downstream calls.
Descriptions prioritize implementation details over LLM guidance. get_rows_fast and get_top_n_fast describe DuckDB, instance_id, row_count thresholds, and file paths, details agents don't control. They lack clear WHEN guidance (e.g., 'Use when user names specific columns and filter operators, not for aggregations or joins'). Descriptions should signal intent, not internal mechanics.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 64 | <=2025-11-25 | v2 |
| 2026-03-09 | D | 59 | - | v1 |
Get the available schema. Returns a JSON object that defines the available subjects (tables) and their columns. Returns a JSON string with: - schemaVersion: int - generatedAt: ISO 8601 UTC timestamp - source: string identifying this server - subjectsCount: int - subjects: [ { name: str, aiDescription: Optional[str], factFieldTypes: { fieldName: { name, type, aiDescription } }, metricFieldTypes: { metricName: { name, type, dimensionsUsed, aiDescription } }, numDimensions: int, numMetrics: int }, ... ]
FAST PATH for rankings and leaderboards. Use when the user asks for "top/bottom N" by a metric (no free-form reasoning). Much faster and cheaper than get_answer. If the output json contains property row_count wil a value less than or equal to 10 then the data property can be used as the answer. If the output json contains a non blank value for the instance_id property then the data property will only contain a sample of the data and the full data set can be found in a table named my_table in a DuckDB database file saved on disk. In that case you MUST use the query_results_fast tool to query the results with SQL to get the data you need to answer the question. This is the only way to access larger datasets.
FAST PATH for querying a result data set produced by other tools. Use when the data set returned by another tool is over a specific number of rows. This gives direct SQL access to the DuckDB database file saved on disk allowing fast and cheap querying of the data. Can be used to access larger datasets without needs further querying to get to an answer. The only table in the database will be named my_table. The columns will be the same as those returned by the tool that produced the dataset. You will have a sample of the data in the data property of the output json from the tool that produced the dataset.
Parameter descriptions are minimal and lack guidance on valid formats/constraints. Examples: 'where' array lacks item schema (what fields/operators are valid?), 'group_by' and 'order_by' lack examples, 'sql' parameter description is one sentence with no hint about table/column naming (only my_table?). This creates ambiguity for LLM tool selection and invocation.
get_financial_periods description is extremely terse (10 chars: 'Get all financial periods available in the data warehouse'). Does not explain WHEN to call it, whether it's a discovery prerequisite, or how results relate to other tools. target_date parameter is optional but lacks context on how it filters results.
Error handling and recovery guidance are weak across most tools. get_rows_fast/get_top_n_fast descriptions mention instance_id behavior but don't explain what to do if query_results_fast fails or how to handle edge cases (empty result, timeout, permission denied). get_answer_slow provides no hints on fallback if the LLM-powered backend fails.
Tool composition is implicit, not explicit. get_rows_fast/get_top_n_fast can return instance_id, which MUST be fed to query_results_fast, but this chain is documented only in description text. No explicit guidance that instance_id is the required chaining ID or what happens if it's omitted. Agents may not infer the required sequence.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All tools are READ_ONLY by domain (data warehouse queries), but this is not declared in the MCP schema. agentic_rag_query may have side effects (caching, logging), unclear if it's safe to retry.