Synthetic Focus Group MCP App - Pneuma

An interactive MCP app for running fast, grounded synthetic focus groups on ads, product concepts, and UI experiences.

Built for the Manufact MCP Apps Hackathon at YC.

Why This Is Useful

• Product teams can test messaging before launch.
• Designers can pressure-test ideas with realistic audience segments.
• Researchers can ask follow-up questions to specific personas or entire subgroups, directly inside the widget.

What We Built

• Two MCP tools:
  • run_synthetic_focus_group
  • ask_persona_followup
• A custom React widget (focus-group-results) that renders:
  • simulation summary
  • persona roster
  • in-widget rerun controls
  • single or batch follow-up chat
• A FastAPI backend that:
  • classifies task type (Objective UI Task vs Subjective Concept)
  • grounds personas with SynthlabsAI/PERSONA from Hugging Face
  • applies sentiment/preference calibration based on actual persona language
  • supports image grounding for visible brand/price cues

How It Works

1. User asks host (Claude/ChatGPT/Inspector) to run a focus group.
2. MCP tool calls backend POST /synthetic/simulate.
3. Backend returns manager summary + personas.
4. MCP server returns widget-linked result.
5. Widget lets the user:
   • edit target audience / stimulus / image and rerun
   • ask follow-ups to selected personas, all personas, or segments

MCP Tools

Tool	Input	Output
run_synthetic_focus_group	target_audience?, stimulus_description, image_url?, persona_count?	Widget props (manager_summary, personas, metadata)
ask_persona_followup	agent_id, question, persona_profile?	Structured answer + supporting memory snippets

Quickstart (Local)

1) Start backend (Terminal A)

cd backend
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
export OPENAI_API_KEY="<your-key>"
uvicorn main:app --host 127.0.0.1 --port 8000 --reload

2) Start MCP server + widget (Terminal B)

npm install
npm run dev

Inspector:

• http://localhost:3000/inspector

Health check:

curl http://127.0.0.1:8000/health

Run in Claude / ChatGPT

Start with tunnel:

npm run dev -- --tunnel

Use the printed URL (example):

• https://<subdomain>.tunnel.mcp-use.com/mcp

Then add that MCP URL:

• Claude: Settings → Integrations → Add integration
• ChatGPT: Settings → Connectors → Add MCP server