MCP App & Agent Use Cases Library

​

The two channels

  TEXT CHANNEL  (≤ 30 lines)          STRUCTURED CHANNEL (structuredContent)
  ────────────────────────────        ──────────────────────────────────────
  What the model reads and can        Metric rows, provenance, freshness,
  quote in its answer.                definitions and quality findings.
                                      What the host draws.

​

The visual surfaces that exist

​

Use cases

​

1. The daily read

• Intent — “How did last week go?”
• Tools — get_daily_report(project_id) for the numbers, then preview_daily_insight(project_id, payload) for the narrative that goes with them. get_daily_insight_readiness first says whether there is anything worth saying.
• What the reader gets — a kpi card, and every figure carries the pull that produced it and how fresh it is.

​

2. Comparing channels

• Intent — “Compare spend efficiency across Meta, TikTok and GA4.”
• Tools — get_daily_report(project_id, connectors=["meta-ads", "tiktok-ads", "google-analytics"]), or a governed question through explore_analyze_query.
• The trap it avoids — CTR, CPA and ROAS are not additive. The semantic ratio views recompute them at the grain you asked for, so a cross-channel comparison is not a sum of ratios. See Agent Rules, rule 2.

​

3. Explaining a drop

• Intent — “Why did conversions fall on 14 July?”
• Tools — get_daily_report for the shape of the dip, search_context and get_context_hub for what the business registered around that date.
• The rule that holds — context events are offered as candidate context, never as causes. toorow records that a deployment and a drop share a date; it does not know that one produced the other.

​

4. Auditing what is connected

• Intent — “Show me every datastream and its mapping status.”
• Tools — flows_list(project_id), then flows_get(project_id, kind, id) for one definition. get_data_quality_report says which of the ten monitors is unhappy and where.
• What the reader gets — the connectors card, and a repair named for each finding rather than a status code.

​

5. Keeping the analysis

• Intent — “Turn this into something I can send.”
• Tools — render_analyze_result freezes each figure, compose_dossier composes them with the narrative into an immutable version, and a Share hands it out — requested by one person, confirmed by a second.
• What the reader gets — a page that draws through the same runtime as the console, with each figure’s provenance beside it. See Renders, Dossiers and Sharing.

​

Next Steps & Cross-References