GlycoCart
A food-ordering assistant for people managing PCOS, prediabetes or glucose on a CGM, built on Swiggy's MCP servers.
Live: https://glycocart.vercel.app
What it does
You sign in with your Swiggy account, answer a short questionnaire, and chat. The agent searches real Swiggy restaurants and dishes for your saved delivery address, estimates how each dish is likely to affect your glucose, and builds a cart with the better options.
- Food delivery (
mcp.swiggy.com/food): dish and restaurant search, full menus, cart, coupons, order history and tracking.
- Groceries (
mcp.swiggy.com/im, Instamart): product search, your usual items, lower-GI swaps.
- Eating out (
mcp.swiggy.com/dineout): restaurants and table availability. Dineout publishes no menus, so GlycoCart gives an ordering brief from the cuisines, never a dish list.
- Daily carb budget from what you have already ordered today, and a retrospective glucose estimate on past orders.
What it will not do
GlycoCart will not place your order, because Swiggy orders cannot be cancelled through the API. Swiggy's docs say to phone support (080-67466729) to cancel. The live Dineout server has no cancel_booking either, so it will not book a table. The agent builds the cart and you check out in the Swiggy app.
This is enforced by the absence of the tools, not by prompt wording: place_food_order, checkout, confirm_order, book_table, create_cart and the payment tools are never registered. npm run test:cart-safety asserts it.
It will not invent a number. Swiggy publishes no per-dish nutrition. GlycoCart matches a dish name against a reference table of common Indian dishes (IFCT-2017 and USDA derived, src/lib/catalog.ts). A dish it cannot recognise is shown as unscored. A category match is labelled a rough estimate. Every glucose figure is an estimate from a closed-form model (src/lib/glycemic.ts) driven by your questionnaire answers, not a clinical measurement and not medical advice.
Stack
- Next.js 15 App Router, React 18, TypeScript strict, Tailwind 3.4
- Vercel AI SDK 4 (
streamText with multi-step tool calling)
@modelcontextprotocol/client with OAuth to Swiggy's MCP servers- Neon Postgres for profiles, chat history and the daily log
- LLM: NVIDIA-hosted models by default on the server key (free). Users can switch to OpenRouter in Settings with their own key; the server never spends an OpenRouter key.
Run locally
npm install
cp .env.local.example .env.local # fill in the values
npm run dev # http://localhost:3000
| Variable | Purpose |
|---|
NVIDIA_API_KEY | Server-side LLM key (nvapi-...). Free tier at build.nvidia.com |
NVIDIA_MODEL, OPENROUTER_MODEL | Optional model overrides |
NEXT_PUBLIC_APP_URL | Public origin. Must match the redirect URI registered with Swiggy |
SWIGGY_TOKEN_SECRET | 32-byte hex. Encrypts session cookies and salts the pseudonymous user id |
DATABASE_URL | Neon Postgres connection string. Tables are created on first use |
Swiggy OAuth redirect URIs:
- Dev:
http://localhost:3000/api/auth/swiggy/callback
- Prod:
https://glycocart.deepansadhukhan.com/api/auth/swiggy/callback
Tests
npm test && npm run build
Sixteen suites, each written against a bug that shipped. Among them: the forbidden-tools check (which also covers any route that names a tool at runtime), the single-restaurant cart guard, WCAG AA contrast in both themes, the rule that the server key is never spent on a paid provider, and the avoid-list and diet matching.
Privacy
- An API key you enter is saved in your browser. The server forwards it to the provider you chose with each request and never stores it.
- Your profile is keyed by a pseudonymous id derived from your Swiggy account, not your name or phone number.
- The LLM sees your questionnaire-derived profile and Swiggy results. There is no CGM integration today.
