POOM MCP App

POOM is an AEO-native marketing intelligence app that turns buyer questions in ChatGPT/Claude into personalized SaaS walkthroughs and qualified vendor leads, by converting Loom, Zoom, or any publicly accessible product video into an interactive POOM experience.

This repository contains the ChatGPT/Manufact MCP app layer (tools + widget UI).

Why this matters

SaaS buyers are increasingly evaluating products inside AI chat, not just search pages.

Buyers ask:

"I'm considering Puzzle. How easy is it to create a journal entry and a PO?"

POOM answers with personalized product proof, not generic marketing pages or long untargeted demos.

The problem

Buyers

• Need quick answers to exact workflow questions.
• Do not want to watch long generic demo videos.
• Need confidence before engaging sales.

Sellers

• Have demo content but limited assistant-native discoverability.
• Need AEO (Answer Engine Optimization), not only SEO.
• Need qualified and enriched lead context, not just form fills.

What this app does

1. Receives a buyer intent query in ChatGPT/Claude.
2. Opens an interactive walkthrough widget for a relevant run.
3. Streams chaptered playback and contextual quiz interactions.
4. Supports structured interaction patterns that can feed lead qualification workflows.

End-to-end in-widget flow

Everything happens inside the ChatGPT widget:

1. Create POOM: user requests a personalized walkthrough from a question or source video.
2. Track POOM progress: widget shows live generation status (ingest, segmentation, script, dubbing, ready).
3. View previous POOMs: user can open past generated POOMs from run history.
4. Play a POOM: user watches the chaptered interactive walkthrough directly in the widget.

No context switching, no external dashboard required.

Demo videos

• POOM demo 1
• POOM demo 2

How the ADK pipeline runs

flowchart LR
  subgraph Chat["Chat Surface"]
    Buyer["Buyer asks workflow question"]
    MCP["POOM MCP App<br/>tools + widget"]
    Buyer --> MCP
  end

  subgraph Runtime["ADK Runtime Pipeline"]
    Ingest["Video ingest<br/>Loom/Zoom/public URL"]
    STT["Audio + STT<br/>FFmpeg + Chirp 3"]
    Gen["Narrative + Production<br/>Gemini + deterministic cut + TTS"]
    Pack["Master artifacts<br/>manifest + chapters + media"]
    Ingest --> STT --> Gen --> Pack
  end

  subgraph Outcome["Interactive Outcome"]
    UI["Chaptered in-chat walkthrough"]
    Lead["Qualified + enriched lead context"]
    UI --> Lead
  end

  MCP --> Ingest
  Pack --> MCP
  MCP --> UI

Loading

What is in this repo

• index.ts: MCP server/tool definitions and backend API bridge.
• resources/tutorial-player/widget.tsx: interactive tutorial player widget.
• resources/tutorial-player/widget.css: widget styling.
• resources/tutorial-player/types.ts: widget prop contracts.
• public/: static assets for app icon/favicon imagery.

MCP tools exposed

• list_runs
• list_pooms
• create_poom
• get_poom_status
• open_run_player
• get_segment_quiz
• submit_segment_quiz

Companion backend dependency

This MCP app expects a runtime API that provides:

• GET /runs
• GET /runs/{run_id}/manifest
• GET /pipeline/jobs
• POST /pipeline/jobs
• GET /pipeline/jobs/{job_id}
• GET /quiz/{run_id}/{segment_id}
• POST /quiz/{run_id}/{segment_id}/score
• media/chapter endpoints returned in manifest payloads

The backend currently used by our team is the POOM/Peazy Explorebook runtime service.

Local development

Prerequisites

• Node.js 22+
• npm
• Running runtime API reachable from this app

Install and run

npm install
npm run dev

Inspector:

http://localhost:3000/inspector

Environment variables

• ADK_API_BASE_URL (default: http://localhost:8000)
• ADK_API_TIMEOUT_MS (default: 9000)
• ADK_DEFAULT_RUN_ID (optional)
• MCP_URL (set by platform on deploy; local default: http://localhost:3000)

Deploy

npm run deploy

Public positioning

POOM helps SaaS teams win assistant-native discovery and conversion by turning existing product videos into personalized interactive walkthroughs and lead intelligence inside ChatGPT/Claude.