Frank Lee: Connecting Claude Code to your analytics stack

What the video covers

Frank Lee is a Principal PM at Amplitude, where he is responsible for agents and MCP products. This episode of the Product Growth Podcast, hosted by Aakash Gupta, was published on February 24, 2026. Lee gives a live demonstration of five workflows built on Claude Code connected to Amplitude’s analytics platform via MCP — and by extension, to ticketing, feedback, and project management tools in the same stack.

The session is focused on a specific problem: the amount of time product managers spend compiling data before they can think about it. Lee calls his approach “vibe PMing” — using agents to handle the analytical heavy lifting so the PM can concentrate on strategy and decision-making rather than data assembly.

Who it’s for

Product managers who do hands-on data analysis, manage analytics platforms, or spend significant time writing weekly reports and synthesizing customer feedback. The demonstrations use Amplitude, Linear, Zendesk, and Gong, so it is most immediately applicable to teams running a similar stack. That said, the underlying workflow logic transfers to other tools once you understand how MCP connections work.

Viewers should be comfortable with basic terminal usage and familiar with at least one analytics platform. This is not an introductory session on Claude Code — it assumes you have already encountered the tool and want to see what a production-grade workflow looks like.

Key takeaways

1. MCP turns Claude Code into a workflow participant, not just an assistant. By connecting analytics, ticketing, and feedback tools through MCP servers, Lee’s agent does not just answer questions about data — it pulls the data, processes it, cross-references multiple sources, and produces structured output in a single run. The distinction matters because it removes the copy-paste step that makes AI tools feel slow in practice.

2. Anomaly detection can run in roughly 90 seconds. When a metric shows an unexpected change on a dashboard, pointing the agent at the chart URL generates a structured root cause investigation. The agent pulls underlying data segments, checks recent product changes, and cross-references customer feedback automatically. A task that previously took three to four hours of manual investigation completes in a single pass.

3. Scheduled agents replace the Sunday report. Lee demonstrates a setup where an agent connects to key dashboards on a schedule, synthesizes the week’s data, flags anomalies, and delivers the summary before the team arrives Monday morning. The agent does not simply extract numbers — it surfaces the patterns and flags what changed, which is typically the hardest part to automate.

4. Feedback synthesis collapses across sources. The agent ingests Zendesk support tickets, Gong call transcripts, NPS survey responses, and app store reviews in a single prompt. It clusters the input by theme, assigns severity to each cluster, and surfaces patterns that would take several days to compile manually. The output feeds directly into triage decisions.

5. From analysis to PRD in the same session. Lee shows how analytics output from Claude Code feeds into a templated product specification draft within Cursor, refined through conversational feedback. The gap between finishing a data investigation and starting a written requirement shrinks to a few minutes.

Worth watching if

You are a PM who regularly spends hours preparing data before meetings or writing up analysis that everyone agrees takes too long. The session is also useful if you are evaluating whether MCP-based tooling is ready for real workflows — Lee’s demonstrations are live, not staged, and the occasional rough edge makes the tradeoffs visible.