Builddy

Builddy — AI App Builder That Turns Tweets Into Deployed Apps

See an app you love? We'll build it for you.

Tweet @builddy Build me a pomodoro timer with dark mode and within minutes, 6 specialized AI agents will plan, design, code, review, polish, and deploy a fully functional web app — then reply with the live link. No sign-up. No coding. Just vibes.

What it does

Builddy is an AI-powered app builder where you describe what you want in plain English (or reply to any tweet with @builddy Build me) and a multi-agent pipeline powered by GLM 5.1 builds a
production-quality web application from scratch. Users watch every step unfold in real time — files streaming into a live code editor as they're generated, pipeline steps updating via SSE, and the
final app deploying instantly with a single click.

How we built it

The 6-Agent Pipeline:

┌──────────────┬────────────────────────────────────────────────────────────┬───────────────────┐
│ Agent │ Role │ Model │
├──────────────┼────────────────────────────────────────────────────────────┼───────────────────┤
│ PM Agent │ Writes a PRD with user stories & acceptance criteria │ GLM-5.1 │
├──────────────┼────────────────────────────────────────────────────────────┼───────────────────┤
│ Designer │ Creates a complete design system (colors, typography, │ GLM-5.1 + Web │
│ Agent │ components) │ Search │
├──────────────┼────────────────────────────────────────────────────────────┼───────────────────┤
│ Coder Agent │ Generates production code using a component library │ GLM-5 │
│ │ │ (streaming) │
├──────────────┼────────────────────────────────────────────────────────────┼───────────────────┤
│ QA Agent │ Validates every acceptance criterion, fixes failures │ GLM-5.1 │
├──────────────┼────────────────────────────────────────────────────────────┼───────────────────┤
│ Polish Agent │ Adds animations, empty states, dark mode, │ GLM-5.1 │
│ │ micro-interactions │ │
├──────────────┼────────────────────────────────────────────────────────────┼───────────────────┤
│ Visual Agent │ Loads the app in Playwright, screenshots it, feeds to │ GLM-5V-Turbo │
│ │ GLM-5V to fix visual bugs │ │
└──────────────┴────────────────────────────────────────────────────────────┴───────────────────┘

The Twitter Loop:
A Playwright-based scraper monitors @builddy mentions every 45 seconds. When someone replies to a
design post with "@builddy Build me", it:

1. Screenshots the parent tweet's content (the app/design being referenced)
2. Captures the original post's text and link card descriptions
3. Feeds everything to the screenshot-to-app pipeline — but instead of cloning, GLM reimagines it
   with a fresh design, added features, and full functionality
4. Deploys the app and replies with the live link via OAuth 1.0a

Fullstack Generation:
For complex requests, Builddy generates complete Express.js + SQLite applications with:

• Auth (JWT + password hashing)
• Database schema with seed data (realistic users, content, activity)
• Multi-page Tailwind CSS frontend with dark mode
• Dockerfile + docker-compose for one-click deployment
• API path rewriting so apps work within Builddy's reverse proxy

Real-Time Streaming:

• SSE (Server-Sent Events) streams every pipeline step to the frontend instantly
• Code appears character-by-character in a live editor as GLM generates it
• Files pop into the file explorer in real time as they're completed
• No polling delay — everything is push-based

Key Features

• Text-to-App — Describe any app, get it deployed in minutes
• Screenshot-to-App — Upload a design screenshot or reply to a tweet, GLM-5V converts it to working code
• Tweet-to-Build — Mention @builddy on X, get a deployed app as a reply
• Live Code Streaming — Watch files generate in real time in the browser
• 6-Agent Pipeline — PM, Designer, Coder, QA, Polish, and Visual agents with distinct roles
• Fullstack Generation — Express + SQLite + Auth + multi-page frontend + Docker
• In-Browser Code Editor — Edit generated files with Cmd+S live save + instant preview refresh
• Gallery + Remix — Browse community-built apps and fork them with your own twist
• Code Autopsy — Paste a dead GitHub repo URL, get a forensic analysis and revival plan
• Light/Dark Mode — Bento-style UI with warm peach light mode and glass-panel dark mode

Challenges we ran into

• Rate limits — GLM-5.1 has a concurrency limit of 2. We built a tiered model strategy: GLM-5.1 for thinking-heavy steps, and GLM-5 for other steps.
• Pipeline stalls — Generating 12 files sequentially with web search + thinking mode caused frequent timeouts. We added per-step timeouts (2-3 min), graceful degradation (skip and continue), and a
  "Force Retry" button that resumes from the exact point of failure.
• File generation quality — Single-shot generation produced 50K+ char monster files. We enforced a
  300-line/12K-char limit, injected a pre-built Tailwind component library, and trimmed context to
  dependencies-only (80% less tokens per file).
• Fullstack deployment — Generated apps need Express running, not just static HTML. We built a
  Process Manager that auto-starts Express servers on dynamic ports, handles npm install, database
  seeding, idle timeouts, and reverse proxies API calls through FastAPI.
• Twitter API paywall — The search/recent endpoint costs $100/month. We bypassed it entirely with a Playwright-based browser scraper that checks the @builddy mentions tab every 45 seconds — free
  forever.

What we learned

• Multi-agent > single-shot. Breaking generation into PM → Designer → Coder → QA → Polish produces
  dramatically better apps than a single "generate everything" prompt.
• Streaming is a game-changer for UX. Watching code appear character-by-character transforms a
  3-minute wait into an engaging experience.
• Context management is everything. Sending 200K chars of