Skip to main content
Capsulate’s MCP connector is hosted and permission-scoped. Read tools return “just enough” detail so your AI client can decide what to fetch next; selected deterministic write tools appear only when you approve mcp:write.

​
Defaults (practical)

  • Search results are capped by limit (use a small value first).
  • Fetch returns a single item by ID (entity, canvas, pattern, message, or thread).
  • Aggregations (records_timeseries, records_aggregate) return logical-record rows plus metadata describing applied filters.
  • Batching: prefer capsulate:bulk_fetch when you need multiple items.

​
Date filters

Date filters accept YYYY-MM-DD or ISO 8601 datetimes. The connector normalizes them and treats date_to as inclusive (end-of-day when you pass a plain date).

​
Messages (privacy + size)

Messages can contain sensitive information. When in doubt:
  • Fetch entities first (structured data), then fetch messages only when the user asks for source context.
  • Prefer snippets/metadata over full bodies if you don’t need the entire content.

​
Performance tips

  • Use capsulate:records_schema to avoid “guessing” fields for grouping/metrics.
  • Keep limit small, then widen only if needed.
  • Use canvases/patterns to scope questions before running aggregations.