Skip to main content

Tableau evidence

Use the capabilities exposed by the current Tableau MCP. Tableau's REST APIs can return workbook/view metadata, view images, and view data, while Metadata API coverage differs by object and permissions. The MCP may expose only a subset.

Official references:

Evidence classes​

  • User-provided: authoritative for intent/preferences, not automatically for current Tableau state.
  • View image: visible title, labels, chart forms, displayed filters/values, annotations, and layout at the rendered state.
  • View/workbook metadata: returned names, project, owner, timestamps, descriptions, tags, and usage only when present.
  • Datasource metadata: returned fields, types, roles, formulas, parameters, logical tables, and lineage only when present.
  • Authorized value query/export: returned values within the selected filters/grain; do not generalize beyond the query.

Resolution​

Treat all IDs and URLs as opaque until the active tool confirms their accepted form. Match by exact stable ID when available, otherwise exact name plus project/owner/site. Ask when ambiguity remains. Follow server pagination for the promised scope.

Visual interpretation​

Inspect the rendered image at sufficient detail. Do not infer hidden filters, actions, tooltips, accessibility conformance, axis truncation beyond visible bounds, or underlying field linkage. A blank render may reflect filters, loading, permissions, or data state—not necessarily a stale extract.

Failure handling​

  • Inventory/authentication failure: explain that live target resolution could not complete.
  • One inaccessible resource: continue with the remaining evidence and identify that coverage gap.
  • Missing image: omit visual claims rather than recreating the layout from metadata.
  • Missing datasource metadata: mark definitions as visual/user-language only.
  • Partial metadata: treat omitted properties as not assessed, not empty.
  • Transient/rate-limit response: follow retry guidance or retry once with bounded backoff.

Place uncertainty where readers use the affected information. Keep diagnostic details concise and sanitize raw errors.