1 of 10

HUMARACART

What v1 Didn't Cover

Coordination over WhatsApp, and how it actually talks to Instamart

2 of 10

THE MESSAGING LAYER

WhatsApp Doesn't Support Group Bots

What You'd Expect

One WhatsApp group.

Everyone in it, talking to one bot together.

What We Built

N separate 1:1 threads with the bot.

The backend maps each one to a single household ID: one shared cart underneath.

The Business API terms forbid group bots, so the household is a backend concept, not a WhatsApp one.

3 of 10

ROLES & PERMISSIONS

Who Can Do What

Member

✓ Add items to the cart

✓ Remove items from the cart

✓ Nudge the holder to check out

Account Holder

✓ Sets the delivery address

✓ Add items to the cart

✓ Remove items from the cart

✓ Confirm checkout & choose payment method

4 of 10

ELI5

What is MCP?

MCP

An open standard that lets an AI agent call external tools in a consistent way, regardless of who built them

Tool

A single callable function a server exposes, with defined inputs and outputs (e.g. search_products)

MCP Client

The agent side. Discovers what tools are available and calls them: this is HumaraCart's role

MCP Server

The provider side. Exposes a set of tools over the protocol: this is what Instamart runs

5 of 10

ARCHITECTURE

Where HumaraCart Sits

HumaraCart

Agent: the MCP Client

Instamart

Backend: the MCP Server

MCP

HumaraCart never talks to Instamart directly: every call and every result crosses this one standard interface.

6 of 10

THE AGENTIC PART

LangGraph + Human-in-the-Loop

agent

calls the LLM, decides next step

tools

calls Instamart MCP

ask_human

pauses the graph for a person

START → agent → tools → agent → … → END

ask_human pauses for two things: picking a variant, or confirming a checkout, whose first call only previews. The real order is placed by a second, code-issued call.

checkout confirmed

START

END

7 of 10

AVAILABLE TOOL CALLS

What HumaraCart Actually Calls

Discover

get_addresses

search_products

Cart

update_cart

get_cart

clear_cart

Payment

get_payment_options

Order

checkout

Track

get_orders

8 of 10

Let's Look at the Docs

mcp.swiggy.com/builders/docs/reference/instamart: the rest of the toolbox, live

9 of 10

CONSTRAINTS & CHALLENGES

What Was Hard

▸ Wrong person could answer a pending question: a race in who a paused question was actually addressed to

▸ Concurrent messages could race on the same session: two messages for one household hitting the agent at once

▸ Sync def, or async, for the background handler?: every message triggers a slow LLM call and an MCP round-trip; both concurrency models were on the table