DeskKit — Agentic Commerce for Razorpay AI Buildathon

Track 1: AI Growth & Agentic Commerce

An autonomous procurement agent that makes a merchant (DeskKit) transactable by an AI buyer, end to end — with every money action explainable, bounded, and gated, and a live audit trail.

What it does

DeskKit is a small office-supplies merchant on Razorpay. This project builds an AI agent that:

• Browses DeskKit's catalog via an MCP tool interface (agent-readable catalog)

• Reasons about purchases and executes real Razorpay test-mode orders

• Enforces a strict monthly budget with two-tier gating: auto-approve under ₹3,000, manual approval required above it, hard block if it would exceed ₹5,000/month

• Persists spend across sessions — the budget genuinely behaves like a monthly limit, not a per-run reset

• Proactively suggests complementary items (upsell/cross-sell) after a successful purchase

• Includes a second AI agent that autonomously generates procurement requests — demonstrating real agent-to-agent commerce, not just human-to-agent

Every decision is traced in LangSmith for a full audit trail, and visualized live in a Monthly Ledger dashboard showing approved/pending/blocked counts and order history.

Related MCP server: razoragent

Why this matches the track

"Every money action explainable, bounded and gated. Show the audit trail and one failure handled gracefully."

This project handles four distinct failure/edge cases gracefully: budget block, approval gate, requests for items not in the catalog, and partial fulfillment of an over-scoped agent-to-agent request — each with clear reasoning and real alternatives, never a silent failure or forced override.

Architecture

Requester Agent (AI buyer) ──> LangGraph ReAct Agent ──> MCP Tool Layer ──> Razorpay Test-Mode API │ │ │ ├── get_catalog │ └── create_order (budget-gated) │ └──> LangSmith (audit trail)

Tech stack

• FastAPI — backend + UI

• LangGraph (ReAct agent pattern) — reasoning and tool orchestration

• MCP (Model Context Protocol) — catalog and checkout exposed as agent-callable tools

• Groq (openai/gpt-oss-120b) — LLM inference

• Razorpay Python SDK — real test-mode payment orders

• LangSmith — observability and audit trail

• Persistent JSON-based spend tracking

Running it locally

python -m venv venv
venv\Scripts\activate          # on Windows
pip install -r requirements.txt

Create a .env file with: RAZORPAY_KEY_ID=your_test_key RAZORPAY_KEY_SECRET=your_test_secret GROQ_API_KEY=your_groq_key LANGSMITH_API_KEY=your_langsmith_key LANGSMITH_TRACING=true LANGSMITH_PROJECT=deskkit-buildathon

Then run:

uvicorn main:app --reload

Visit http://127.0.0.1:8000.

Demo

Type an instruction (e.g. "Buy 10 USB-C cables and 1 laptop stand"), or click Simulate Requester Agent to see a second AI generate its own purchase request — true agent-to-agent commerce.

Screenshots

Successful order with real-time budget tracking

Budget gate blocking an over-limit order, with graceful reasoning

Monthly ledger — full audit trail at a glance

Agent-to-agent commerce: a requester agent generates its own purchase request

Built for the Razorpay AI Buildathon, Track 1.