A small proof of concept for browsing recent arXiv papers and turning one selected paper into a podcast-style script.

• Browse recent arXiv papers across focused research areas.
• Generate an 8-turn grounded dialogue from a selected paper.
• Name both speakers and choose a separate OpenAI model for each one.
• Replay the generated script with live dialogue animation.
• Export the script as plain text or ElevenLabs-compatible JSON.

This version intentionally has no database persistence.

• Paper discovery is fetched live from arXiv.
• Search results use the paper abstract directly.
• Script generation happens on demand through a local Node API and the OpenAI Responses API.
• Nothing is stored. Download the generated script if you want to keep it.

• Frontend: React 18, TypeScript, Vite
• Styling: Tailwind CSS, Radix UI primitives
• API: Express
• AI: OpenAI

nvm use
npm install
cp .env.example .env
# fill in OPENAI_API_KEY
npm run dev

Frontend: http://localhost:8080 API: http://localhost:3001

npm run dev
npm run build
npm run preview
npm run lint
npm run check
npm run healthcheck

• Node 20 is the expected local and CI runtime.
• .nvmrc is included for local version alignment.
• GitHub Actions validates dependency install, audit, mock script evaluation, lint, typecheck, build, and a server health smoke test.

HOST=127.0.0.1
OPENAI_API_KEY=
OPENAI_SCRIPT_MODEL=gpt-5.5
OPENAI_SCRIPT_MAX_TOKENS=1600
OPENAI_SCRIPT_TURN_MAX_TOKENS=1200
PORT=3001
ARXIV_USER_AGENT=

Only OPENAI_API_KEY is required. For local-only secrets, .env.local is also loaded by the Node API and can contain the same keys. The UI lets each speaker choose between gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5-mini, and gpt-5-nano before script generation. ARXIV_USER_AGENT is optional and can be used to identify your deployment when calling arXiv RSS.

• npm run dev starts both the Vite frontend and the local API server.
• npm run preview builds the frontend and serves the built app through the API server.
• npm run healthcheck hits GET /api/health on the configured local port.
• The frontend proxies /api/* requests to the local server during development.

This repo is optimized for lightweight POC hosting, not a full production platform.

• Build the frontend with npm run build.
• Run the server with npm run start.
• Set OPENAI_API_KEY in the runtime environment.
• Serve the repository as a single Node process. The Express server will serve dist/ automatically after a build.
• Serverless-style /api/* entrypoints are included for Vercel-compatible hosting.

Good fits:

• Vercel
• Railway
• Render
• Fly.io
• A small VPS with Node 20+

Not included on purpose:

• Docker
• database migrations
• background workers
• secrets management beyond env vars

src/
  components/   React UI
  hooks/        Frontend search, selection, and script hooks
  pages/        App routes
  constants/    Research area definitions

server/
  index.ts      Express API entrypoint
  research.ts   arXiv search and filtering
  openai.ts     Podcast script generation
  types.ts      Shared server-side types

• No auth
• No paper history
• No episode library
• No background jobs
• No database

GET /api/health returns:

{"ok":true}

MIT