Skip to main content
Data MCP turns conversation data into something you can interrogate in plain language. The flow below goes from “what can I even filter on?” to a single transcript, using the tools in the order you’d naturally reach for them.
You don’t call these tools by hand — describe what you want and let the client chain the tool calls. Example prompts are shown throughout. Every tool takes an optional project_id; scope your questions to a project when your account spans several.

​
1. Discover what you can filter on

Start with get-metrics to see the metrics available for your account — the built-in ones plus any custom metrics your projects define. This tells you what you can filter and sort conversations by.
“What metrics can I filter conversations on for project acme-support?“

​
2. Find the conversations you care about

Use search-conversations to filter on those metrics over a date range, then sort and page through the results. Combine filters with and/or and sort by start time, duration, or ID.
“Find last week’s conversations with a poly_score below 3, longest first.”

​
3. Search across transcripts

When you’re chasing a phrase rather than a metric, search-transcripts does full-text search over transcript turns and returns each match with surrounding turns for context.
“Search transcripts for ‘cancel my policy’ in the last 30 days and show a couple of turns either side.”

​
4. Read a full conversation

Once you’ve found a conversation worth reading, get-conversation returns every turn, top to bottom.
“Show me the full transcript for conversation abc-123.”

​
Why this matters

Each tool is small, but together they cover the loop analysts actually run: what can I measure → which conversations fit → where does the phrase show up → what happened in this one. Because it all runs through your MCP client, you can follow a thread — from an aggregate concern to a single call — without leaving your chat.
Keep results tight — filter to the conversations that matter and scope to a project_id — so the model only ingests what the task needs. See Security & safe use.
Last modified on July 17, 2026