Meeting prep

Input
Output

🚀 Inspiration

We’re all guilty of scrambling for context five minutes before a call—jumping between LinkedIn, company “About” pages, and half-baked Google searches. Meeting prep turns that chaos into a single click: paste a LinkedIn URL and instantly receive an ice-breaker plus smart, open-ended questions you can lead with. No more tab-surfing, no more awkward silences.

🪄 What it does

One-field input. The React front-end (frontend/src/App.tsx) takes a LinkedIn URL and optional notes.
Streaming magic. A single POST to /api/report-stream opens a Server-Sent Events (SSE) channel. While Claude and Perplexity are researching, we push live logs—tool calls, intermediate thoughts, even spinner animations.
Multi-tool reasoning.
- Perplexity Ask (via @chatmcp/server-perplexity-ask) handles real-time web look-ups.
- Claude 3 Sonnet synthesizes findings into a strict JSON dossier via the Model Context Protocol (MCP).

Zero parsing errors. The custom MCP tool return_meeting_dossier forces Claude to respond with:

{
 "status": "complete",
 "opener": "…",
 "questions": [ { "q": "…", "why": "…" } ]
}

Front-end render. As soon as status === "complete", React swaps the spinner for:

a two-sentence personalized ice-breaker
3–5 high-leverage questions, each with a one-line rationale.

Total turnaround: 7–12 s, including real-time progress.

🏗️ How we built it

Layer	Key files / libs	Highlights
Front-end	`App.tsx`, `index.css`, React 18, TypeScript	SSE reader decodes incremental chunks, progressive logs with `fadeIn` CSS, mobile-first monospace UI
API	`backend/server.js`, Express, Zod, helmet, rate-limit	Two endpoints: JSON (`/api/report`) and streaming (`/api/report-stream`) with abuse protection (30 req/min)
AI orchestration	`backend/mcpClient.js`, `@modelcontextprotocol/sdk`, Anthropic SDK	Launches Perplexity Ask as a local MCP tool server; dialogue loop retries until `return_meeting_dossier` fires; schema-validated JSON
Security & Ops	`.env.example`, CORS, TLS-ready reverse proxy	No secrets in repo; keys via `process.env`; helmet removes `X-Powered-By`

🧱 Challenges we ran into

Gumloop scraping limits. Free tier throttled to 100 req/day → added caching and graceful fallback to user-pasted notes.
Strict JSON or bust. Claude loves to chat; MCP tools force discipline, but only after we sandboxed the convo and validated each reply.
SSE in React 18. Vanilla EventSource has no back-pressure; we built a getReader loop to stream >64 KB payloads without freezes.

✨ Accomplishments we’re proud of

Full-stack streaming in <48 h: URL → live logs → final dossier.
99.9 % JSON validity after schema-driven retries—zero judge-side parsing errors.
Security-first defaults: helmet, rate-limit, env-only API keys, zero-trust CORS.
UX delight: retro monospace, playful 2 px borders, micro-animations.

📚 What we learned

Model Context Protocol is the glue that lets any LLM call any tool—no bespoke adapters.
Prompt engineering ≠ copywriting: enforcing structure matters more than adjective tuning.
SSE > WebSockets for one-way, fire-and-forget updates—lighter infra, no ping/pong.

🔮 What’s next for Meeting prep

Milestone	Description
Calendar plug-in	Auto-detect participants from invites → pre-generate dossiers 5 min before the call.
Slack slash-command	`/meetingprep @alice` returns dossier in-thread.
Multi-party fusion	Collate profiles of all attendees and suggest cross-team synergy questions.
Voice mode	Whisper transcription of intros → real-time follow-ups in EarPods.

🛠️ Built With

Backend: Node.js • Express • Zod • express-rate-limit • helmet • dotenv • picocolors
AI / Tooling: Anthropic Claude 3 Sonnet • Perplexity Ask • Model Context Protocol SDK
Frontend: React 18 • TypeScript • Vite/CRA • vanilla CSS animations
Utilities: Gumloop Scrape API • AbortController • Server-Sent Events

▶️ Try it out

git clone https://github.com/your-org/when2meet-mcp-2025.git
cd when2meet-mcp-2025
cp backend/.env.example backend/.env    # add your API keys
npm run dev                             # starts both front- and back-end
# open http://localhost:5173 and paste any LinkedIn URL

(Public demo will be live at *https://meetingprep.ai** once quotas reset.)*