Repromo

Inspiration

Builders ship apps every day. Most of them still have no demo video.

You vibe-code a product over a weekend, deploy it, then hit the same wall: writing a pitch, storyboarding shots, and rendering clips takes longer than building the app. So launches go out with screenshots, a Notion doc, or silence.

We wanted a showrunner that meets builders where they already are — a website or a GitHub repo — and turns that into a short demo they can post the same day. Not a slide about AI. A real video.

What it does

Repromo turns a public website or GitHub URL into a short demo / launch video.

1. Paste a link
   
2. Agents read what you built, write a pitch, plan scenes, then film clips
   
3. Watch progress live, pause / stop if needed, then download the MP4
   

Under the hood it uses Qwen for creative work and HappyHorse for text-to-video on Alibaba Cloud DashScope / Qwen Cloud (Token Plan). No mock videos.

How we built it

• Frontend: Next.js App Router, HeroUI, light UI with brand logo, favicons for pasted domains, download buttons
  
• Orchestration: LangGraph.js showrunner — parse → scout → script → storyboard → generate_shots → finalize
  
• LLM: Qwen (qwen3.7-plus) via Token Plan OpenAI-compatible endpoint
  
• Video: HappyHorse (happyhorse-1.1-t2v) async synthesis + polling
  
• Jobs: Neon Postgres so status survives Vercel serverless
  
• Safety: IP rate limits on generate / download / job control
  

Alibaba Cloud proof in-repo:

• src/lib/qwen/client.ts
  
• src/lib/video/happyhorse.ts
  

Architecture: docs/ARCHITECTURE.md

Challenges we ran into

• Serverless job store — in-memory jobs 404’d across Vercel instances; we moved to Neon
  
• Billing confusion — free-tier vs Token Plan keys/endpoints are not interchangeable; we wired the Token Plan base URL and key
  
• Long video renders — HappyHorse is async and slow; we added polling UI, pause/stop, and clear errors
  
• Cross-origin downloads — DashScope video URLs needed a download proxy with host allowlisting
  
• Abuse risk — cheap to spam generates; we added per-IP rate limits
  

Accomplishments that we're proud of

• A full AI showrunner path from link → pitch → storyboard → real HappyHorse clips
  
• Honest product — fail-fast, no fake videos, errors surface in the UI
  
• Hackathon-ready packaging: public MIT repo, Alibaba proof links, architecture diagram, live deploy
  
• Polish that matters for demos: logo PNGs, pause/stop, download, rate limits
  

What we learned

• Agent pipelines are only as trustworthy as job state and billing wiring
  
• Qwen Cloud Token Plan, Coding Plan, and pay-as-you-go are separate credentials and hosts
  
• For Track 2, fewer high-quality shots beat a long fragile pipeline
  
• Judges and users need the same thing: a clear progress story and a file they can take away
  

What's next for Repromo

• Stitch multi-shot clips into one timeline with music / VO
  
• Website screenshot → image-to-video for more product-faithful scenes
  
• Brand kit upload (logo, colors, CTA) for on-brand outputs
  
• Durable queue (beyond after()) for longer HappyHorse runs
  
• Team workspaces and shareable public demo pages
  

Built With

• nextjs
• qwencloud