Watch a five-agent AI crew turn a one-line idea into a working proof of concept — live, in your browser.
Live demo: Agent Crew Studio
This is a fun experiment, not a production tool. The goal is simple: give a crew of AI agents — modeled loosely on how a small engineering team actually works — a one-line product idea, and watch what they do with it. Each agent gets a role (product, design, backend, frontend, test) and hands its output to the next, the same way a PRD moves from a product owner to an engineer to a build.
The bar for the output isn't "production-ready." It's "can this idea's core value be demoed and validated" — the smallest thing that proves the concept works. The code the agents write can have bugs. It is not run, executed, or tested by anything other than the agents' own (unverified) judgment — so treat whatever comes out as a rough first draft, not something to ship.
Five agents run in a fixed sequence, each handing its output to the next:
| Agent | Produces | Does |
|---|---|---|
| Product Owner | prd.md |
Turns your idea into a lean PRD: users, flows, features, and — just as important — what's deliberately left out. |
| Engineering Lead | design.md |
Turns the PRD into a design: backend module structure and function signatures, the exact HTTP API contract between backend and frontend, and the frontend's page structure. |
| Backend Engineer | backend.py |
Implements the design in pure Python (standard library only), exposed over a small HTTP server so the frontend can call it for real. |
| Frontend Engineer | index.html |
A single static HTML page (vanilla JS, one CSS stylesheet, no build step) that calls the real backend API. |
| Test Engineer | test_*.py, README.md |
Reviews the backend code against the API contract (and fixes clear bugs it finds — the last agent to touch the code before it's yours), writes unit tests, and writes a short README explaining how to run everything. |
All five artifacts stream into the browser live as they're written — the flow diagram and output panel update in real time — and are downloadable as a zip when the run finishes.
- Backend: Python 3.13, CrewAI 1.14.7, FastAPI + uvicorn
- Frontend: Next.js 16 (App Router), React 19, TypeScript, Tailwind, @xyflow/react
- Model provider: OpenAI only, bring-your-own-key
# backend
cd backend
uv run uvicorn app.main:app --reload # http://localhost:8000
# frontend, in another terminal
cd frontend
NEXT_PUBLIC_API_URL=http://localhost:8000 npm run dev # http://localhost:3000(Or put NEXT_PUBLIC_API_URL=http://localhost:8000 in frontend/.env.local
so you don't have to set it inline each time.) Open localhost:3000, paste
an OpenAI API key via the header button, and submit an idea.
The key you paste in is held in browser memory only, for that tab — it is
sent with your run request and never written to disk, localStorage,
cookies, or logs, on either side. Refreshing the page clears it. No
account, no login, nothing saved.
- No auth, no accounts, no saved history.
- No persistence — everything lives in server memory for the life of the process; a restart loses every in-progress and completed run.
- One run at a time per browser session; a hard cap of 3 concurrent runs server-wide.
- Runs time out after 15 minutes.
- Generated code is never executed or tested by the studio itself, on either side of a run — the crew's own read-through review is the only check any of it gets.
- Per-agent model selection isn't exposed in the UI — it's fixed in the crew's own config.