Read-only MCP server for exploring and analysing a Secoda data catalog
Secoda Analysis MCP server demonstrates solid definition quality with consistent naming conventions, comprehensive descriptions, and well-structured input schemas across all 12 tools. All tools follow verb_noun naming (search_, retrieve_, list_, get_, ai_chat). Descriptions are detailed and contextual (range: 60-350 chars, well within 10-1024 baseline). Parameter schemas are complete with type declarations and descriptions. However, output schemas are not documented, responses are referenced in descriptions but formal output structures are not declared. Error handling is present but generic (relies on API responses rather than tool-level guidance). No tool annotations (readOnlyHint) despite all being read-only operations.
Start an AI chat session in Secoda and wait for the response. Submits a prompt to the Secoda embedded AI endpoint and polls until the response is complete. Sends MCP progress notifications at each poll interval so clients can show elapsed time. Returns the AI's response text along with the chat ID, which can be passed as `parent` in a follow-up call to continue the conversation.
Retrieve the upstream and downstream lineage of an entity.
Retrieve a specific collection by its ID. Use this to get full details about a collection after finding it via list_collections.
Retrieve a specific question and its answer by ID. Use this to read the full content of a question and any accepted answers after finding it via list_questions or search_documentation.
Retrieve a specific resource by its ID. Use this to get full details of a catalog resource (table, column, view, etc.) after finding it through list_resources or search_data_assets.
Retrieve all business term definitions from the workspace glossary.
Output schemas not documented. Tool descriptions reference what results are returned but formal output structure (fields, types) is not declared. LLMs cannot reliably extract fields from downstream responses or plan multi-step chains.
No tool annotations despite all 12 tools being read-only. readOnlyHint would signal to MCP clients that these operations are safe to retry and side-effect-free.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 13 | 2026-07-28+ | v2 |
| 2026-03-09 | D | 52 | - | v1 |
List all collections in the workspace. Collections are organized groups of related resources (tables, dashboards, documents). Use this to browse what topic areas and resource groups exist.
List all questions in the workspace. Questions in Secoda represent data consumer inquiries that have been asked and answered by the data team. Use this to find existing answers before asking a new question.
List catalog resources with advanced filtering capabilities. This endpoint provides precise control over resource queries using structured filters. Use this when you need exact matching, specific field filtering, or complex queries.
Retrieve an entity from the catalog by ID.
Search for data assets in the catalog.
Search for documentation in the catalog.
Error handling is generic. Tool responses delegate to the Secoda API client, which returns raw errors or JSON without recovery guidance. No tool-level error classification (retryable, user-fixable, fatal) or actionable recovery hints.
Parameter 'truncate_length' appears in 11 of 12 tools with minimum=1 constraint. This numeric parameter lacks a reasonable maximum bound, allowing LLMs to pass absurdly large values (e.g., 999999999). Should define maximum (e.g., 10000).
glossary() tool accepts only truncate_length parameter. Description is generic (60 chars) and does not explain when to call this vs search_documentation for glossary terms. Lacks guidance on output structure (is it a dict, array, what fields?).