swiggy-party-agent

Your team fills a Google Form. The agent reads it, builds a Swiggy cart, and waits. You confirm. The order is placed.

Built on Swiggy Builders Club MCP + Vercel AI SDK + Anthropic.

---

How it works

Google Form → CSV export → party-agent → Swiggy MCP → Cart → Confirm → Order placed

1. Share a Google Form: name, dietary restrictions, cuisine preferences, spice level.
2. Team fills it in. Export responses as CSV from Google Sheets.
3. Run the agent (CLI, Slack bot, web app, or MCP server — pick your surface). The agent:
   • Resolves the office delivery address via get_addresses
   • Finds a restaurant that fits the group via search_restaurants
   • Picks a dish per person via search_menu
   • Builds the full cart in one update_food_cart call
   • Applies the best available COD coupon
4. The agent presents a per-person breakdown. You confirm.
5. Agent calls place_food_order.

The agent never places an order on its own. Cart building and order placement are separate steps.

---

Integration surfaces

The core agent logic lives in src/agent.ts and is surface-agnostic. The same buildCartForGroup and placeOrder functions work across every integration below.

CLI (included)

Run it from a terminal. Best for one-off events or testing.

npm run dev -- --csv responses.csv --event "v2.0 Launch" --address "Office" --budget 250

Slack bot

Wire the agent to a Slack slash command or workflow. The team lead runs /party-order in the team channel, the bot posts the cart summary as a message with Approve / Skip buttons, and calls placeOrder on approval.

/party-order event="v2.0 Launch" csv=<Google Sheets URL>

Suggested stack: Slack Bolt (Node.js) + Block Kit for the confirmation UI. The agent reads the sheet via Google Sheets API instead of a local CSV export.

Web app

A Next.js app where the team lead pastes a Google Sheets link, reviews the AI-built cart in a table UI, edits individual items if needed, and clicks Place Order. Each team member's row is editable before confirmation.

Suggested stack: Next.js + Tailwind + Swiggy MCP on the server side.

MCP server (expose as a tool)

Expose buildPartyCart and placePartyOrder as MCP tools so any MCP-compatible AI client (Claude Desktop, Cursor, etc.) can trigger a team order from a conversation.

User: "Order lunch for the team. Here's the sheet: <url>"
Agent: [calls buildPartyCart tool] → presents summary → [calls placePartyOrder on confirmation]

Suggested stack: @modelcontextprotocol/sdk server, two tools, same agent logic underneath.

Scheduled (cron)

Run every Friday at noon. Pull the latest sheet responses, build the cart, post the summary to Slack for approval. If no one rejects within 30 minutes, place the order automatically.

Suggested stack: GitHub Actions cron + Slack incoming webhook for the approval window.

---

Architecture

┌──────────────────────────────────────────────────────────────────┐
│                        Integration Surface                       │
│         CLI  │  Slack Bot  │  Web App  │  MCP Server  │  Cron    │
└──────────────────────────────┬───────────────────────────────────┘
                               │
               ┌───────────────▼───────────────┐
               │          src/agent.ts         │
               │                               │
               │  buildCartForGroup()          │
               │  placeOrder()                 │
               └───────────────┬───────────────┘
                               │
       ┌───────────────────────┼───────────────────────┐
       │                       │                       │
┌──────▼──────┐        ┌───────▼───────┐     ┌────────▼────────┐
│ src/parser  │        │ Vercel AI SDK │     │  Swiggy Food    │
│             │        │ + Anthropic   │     │  MCP Server     │
│ CSV / Sheets│        │ (any model)   │     │                 │
│ → TeamMember│        │               │     │ get_addresses   │
│ []          │        │               │     │ search_*        │
└─────────────┘        └───────────────┘     │ update_cart     │
                                             │ place_order     │
                                             └─────────────────┘

Cart cap

Defaults to ₹5000 per order, configurable via --cap (CLI) or the cartCap parameter (programmatic). Large teams split into groups automatically — each group gets its own cart and confirmation step.

Swiggy Builders Club v1 sandbox accounts have a ₹1000 hard cap. Pass --cap 1000 if you're on a sandbox client_id.

---

Setup

Prerequisites

• Node.js 20+
• Anthropic API key
• Swiggy account with a saved delivery address labelled "Office"
• Swiggy Builders Club access token (apply, or get one via the consumer quickstart)

Install

git clone https://github.com/Siddhant-K-code/swiggy-party-agent.git
cd swiggy-party-agent
npm install

Configure

cp .env.example .env

Edit .env:

ANTHROPIC_API_KEY=sk-ant-...
SWIGGY_ACCESS_TOKEN=eyJhbGci...   # from Swiggy OAuth flow

# Optional — any Anthropic model ID. Defaults to claude-opus-4-5.
# ANTHROPIC_MODEL=claude-3-5-haiku-20241022

Get a Swiggy access token

Swiggy MCP uses OAuth 2.1 with PKCE. Quickest path for testing:

1. Add Swiggy MCP to Claude Desktop via the consumer config
2. Complete the OAuth flow (phone + OTP)
3. Extract the token from Claude Desktop's MCP config, or run the PKCE flow directly via the developer quickstart

Tokens last 5 days. Re-run the OAuth flow when one expires.

---

CLI usage

1. Create the Google Form

Use the template in examples/google-form-template.md.

Fields: Name, Dietary Restrictions, Cuisine Preferences, Dish Preferences, Spice Level.

2. Export responses as CSV

In Google Sheets: File → Download → Comma Separated Values (.csv)

See examples/responses.csv for the expected format.

3. Run

npm run dev -- --csv responses.csv --event "v2.0 Launch Party" --address "Office" --budget 250

Or after building:

npm run build
node dist/cli.js --csv responses.csv --event "v2.0 Launch Party" --address "Office" --budget 250

Options:

Flag	Default	Description
--csv	(required)	Path to the exported Google Form CSV
--event	"Team Party"	Event name shown in the summary
--address	"Office"	Label of the saved Swiggy delivery address
--budget	250	Max spend per person in INR (used to size groups)
--cap	5000	Max cart total per order in INR

4. Confirm and place

  Order 1 of 2 — Biryani House
  ─────────────────────────────────────────
  Rahul Sharma      Chicken Biryani (Half)  ₹249
  Ankit Gupta       Chicken Fried Rice      ₹199
  Vikram Singh      Butter Chicken + Naan   ₹329
  ─────────────────────────────────────────
  Coupon  : LAUNCH20  -₹50
  Total   : ₹727

? Place order 1/2 from Biryani House (₹727)? (y/N)

y places the order. n skips that group.

---

Programmatic usage

Import the agent directly to build any integration:

import { buildCartForGroup, placeOrder } from "./src/agent.js";
import { parseCSV, splitIntoGroups } from "./src/parser.js";

const members = parseCSV("responses.csv");
const groups = splitIntoGroups(members, 250, 5000);

for (let i = 0; i < groups.length; i++) {
  const summary = await buildCartForGroup(
    groups[i], "Office", 250, i, groups.length, 5000
  );

  // present summary via your surface (Slack, web, etc.)
  // then on confirmation:
  const orderId = await placeOrder(summary, 5000);
}

---

Dietary restrictions

The parser normalises free-text entries to canonical tags:

Input	Tag
Vegetarian, Veg	vegetarian
Vegan	vegan
Jain	jain
Halal	halal
Gluten free, Gluten-free	gluten-free
No peanuts, Peanut allergy	no-peanuts
No dairy, Lactose	no-dairy
Eggetarian	eggetarian

Tags are enforced in the agent prompt. A member tagged vegetarian only receives veg items.

---

Limitations

• COD only: Swiggy MCP v1 supports Cash on Delivery. Online payment coupons are filtered out.
• Single restaurant per group: Each cart is tied to one restaurant. The agent picks the best fit.
• Token expiry: Access tokens last 5 days. No automatic refresh in v1.
• India only: Addresses must be in a Swiggy-serviceable city.

---

Built with

• Swiggy Builders Club — Food MCP server
• Vercel AI SDK — MCP client + generateText with tool use
• Anthropic Claude — configurable via ANTHROPIC_MODEL, defaults to claude-opus-4-5
• csv-parse — CSV parsing
• inquirer — confirmation prompts

---

Author

Siddhant Khare (@Siddhant_K_code)