Generate a personalised multi-day meal plan, browse full recipes from real cooking sites, track what's in your pantry, and turn the plan into a shopping checklist with one-click search links to Blinkit, BigBasket and Instamart.
Live demo: https://forkcast-frontend.onrender.com
No account needed. Open the demo, answer a few questions about diet, budget and time, and get a full multi-day plan with recipes. A Demo Mode with a sample plan is built in, so the app works even without an AI key.
- Personalised meal plans — multi-day plans generated for your diet, cuisine, budget and cooking time, streamed live as they are built.
- Real recipes — every meal links to a full recipe from a real cooking site, not a generated summary.
- Two generation modes — AI Generated (pure model output) and Web Search (recipes found on the live web via Tavily).
- Pantry tracking — keep a list of what you already have, and the planner cooks around it.
- Shopping checklist — turn a plan into a checklist with one-click search links to Blinkit, BigBasket and Instamart.
- Demo Mode — a sample plan ships with the app, so visitors can try everything without waiting on the AI.
The frontend talks directly to the Express backend over SSE (server-sent events), so the plan streams in day by day instead of making you wait for one big response. There is deliberately no proxy in front of the API: Render's static-site rewrites buffer streaming responses, and the direct call keeps the stream unbuffered.
The AI chain is Gemini first, Groq second. The default model is gemini-3.5-flash-lite, which is roughly 95 percent faster per day than the full flash model (about 4 to 9 seconds a day versus 90 to 170). If Gemini is overloaded or fails, the request falls back to Groq's openai/gpt-oss-20b automatically.
Every AI response is validated with Zod before it reaches the frontend, and recipe images go through an SSRF-safe proxy, so the backend never fetches an arbitrary URL the model invented.
| Layer | Tech |
|---|---|
| Frontend | React 19, TypeScript, Vite, Zustand |
| Backend | Node.js, Express 5, TypeScript |
| AI | Google Gemini (gemini-3.5-flash-lite), Groq (openai/gpt-oss-20b), Tavily web search |
| Validation | Zod |
| Tests | Vitest (52 tests across frontend and backend) |
| Deploy | Render Blueprint (render.yaml): Express web service + Vite static site |
forkcast/
frontend/ React + Vite app (Zustand state, SSE plan streaming)
backend/ Express 5 API (plan generation, recipes, pantry, rate limiting)
render.yaml Render Blueprint for both services
DEPLOY.md Full deployment steps
# Backend (needs GEMINI_API_KEY, and GROQ_API_KEY for the fallback)
cd backend
npm install
npm run dev
# Frontend
cd frontend
npm install
npm run devThe backend runs on http://localhost:3000 and the frontend on Vite's dev server. Set VITE_API_URL to the backend URL if it differs.
| Variable | Where | Required |
|---|---|---|
GEMINI_API_KEY |
Backend | Yes, for AI generation |
GROQ_API_KEY |
Backend | For the fallback model |
TAVILY_API_KEY |
Backend | For Web Search mode only |
VITE_API_URL |
Frontend | Yes, full backend URL |
CORS_ORIGIN |
Backend | Yes, full frontend URL |
Secrets are set in the Render dashboard, never in render.yaml. Note that Render's fromService host returns just the service name, not a URL — always use the full https:// URL for VITE_API_URL and CORS_ORIGIN.
This repo ships a Render Blueprint. In the Render dashboard: New → Blueprint → connect dropps07/forkcast → Apply. See DEPLOY.md for the full steps.
Both services run on Render's free tier, so they sleep when idle — the first visit after a while can take a minute to wake up.
Generation is capped to protect the shared Gemini key (free tier: 20 requests a day per model across all visitors):
- Anonymous: 5 generations per IP per day
- With an access code: 10 per day
- Development: 1000 per day
Counts are stored in memory on the backend, so a redeploy resets them.
- A database backend, so plans, pantry and preferences persist across devices and redeploys (currently everything lives in the browser and in-memory server state). Accounts stay optional, generating a plan never needs a login.
- Saved and favourite recipes.
- Plan history and sharing.
MIT