Apply for AI Grants India

WebMCP is an emerging approach for making website capabilities available to AI agents through structured, permissioned tools rather than relying only on visual browsing. In Indian e-commerce, this creates an important opportunity: connect conversational AI shopping experiences with the Open Network for Digital Commerce (ONDC) while preserving merchant control, buyer consent, protocol compliance and transaction security.

The practical question is not whether WebMCP replaces ONDC. It does not. ONDC provides the network and protocol framework for interoperable commerce, while WebMCP can act as an agent-facing interface that helps AI systems discover and invoke approved commerce actions. Together, they can enable an AI assistant to find products, compare sellers, verify delivery eligibility, create carts and initiate checkout through ONDC-compatible flows.

What Is WebMCP?

WebMCP can be understood as a model context protocol layer exposed by a website or web application. It describes useful actions, their inputs, outputs, permissions and constraints in a machine-readable way. An AI agent can then call these actions instead of attempting to interpret every page element through screenshots, DOM inspection or fragile browser automation.

A fashion marketplace, for example, might expose tools such as:

• search_products
• get_product_details
• check_serviceability
• create_cart
• apply_coupon
• get_delivery_quote
• initiate_checkout
• track_order
• request_return

A well-designed tool should return structured data, validate inputs on the server and clearly identify any action that creates financial, legal or fulfilment consequences. WebMCP should therefore be treated as an agent interaction layer, not as an alternative to authentication, authorisation, payment security or commerce protocols.

What Are ONDC Protocols?

ONDC is an open, interoperable network intended to separate buyer and seller applications from the underlying commerce network. Instead of forcing consumers and merchants to use one closed marketplace, ONDC enables participants such as buyer network participants, seller network participants, logistics providers and other service providers to communicate using defined APIs and protocol conventions.

ONDC transactions commonly involve a sequence such as:

1. A buyer application creates a search request.
2. Seller applications return catalogues or offers.
3. The buyer selects an item and requests an order quotation.
4. The buyer confirms billing, fulfilment and payment information.
5. The order is created and acknowledged.
6. Updates are exchanged for fulfilment, delivery, cancellation, returns and support.

The exact fields, message formats, version requirements and participant responsibilities depend on the relevant ONDC domain and implementation specification. Any production integration must therefore use the applicable current ONDC documentation and validation process rather than assuming that a generic API wrapper is sufficient.

How WebMCP Can Be Used in Indian E-commerce to Integrate with ONDC Protocols

The strongest architecture places WebMCP above an ONDC integration service. The AI agent interacts with narrowly scoped WebMCP tools, while the backend translates those requests into valid ONDC messages and manages responses from network participants.

A simplified flow looks like this:

AI assistant
    ↓
WebMCP tool layer
    ↓
Agent gateway and policy engine
    ↓
ONDC adapter / buyer or seller network integration
    ↓
ONDC network participants
    ↓
Seller, logistics and payment systems

This separation is important. An AI model should not directly construct and transmit arbitrary ONDC requests with unrestricted credentials. Instead, the gateway should enforce schemas, identity, consent, rate limits, business rules and transaction state transitions.

1. AI-powered product discovery

A shopper could ask, “Find a vegetarian meal kit under ₹800 delivered to Bengaluru by tomorrow.” The AI agent can call a WebMCP search tool with structured parameters such as category, dietary attributes, maximum price, pincode, delivery deadline and quantity.

The backend can convert those parameters into an ONDC-compatible search request, query relevant seller participants and normalise responses into a common product model. The agent can then compare seller offers using factors such as:

• Total payable amount, including delivery and taxes where available
• Seller name and fulfilment information
• Estimated delivery window
• Return or cancellation terms
• Product attributes and availability
• Ratings or other permitted trust indicators

The normalisation layer must preserve important distinctions. It should not merge two offers merely because their product names are similar, and it should clearly show when delivery estimates or fees are conditional.

2. Seller catalogue access

For seller-side applications, WebMCP can expose tools that help merchants or their authorised agents manage ONDC catalogues. Examples include validating a product listing, updating stock, changing prices, mapping categories and checking missing mandatory attributes.

An agent could identify products with incomplete images or inconsistent inventory, but changes should require explicit permission and strong audit logging. Bulk edits should have preview and rollback capabilities, particularly for pricing, stock and tax-related fields.

3. Serviceability and delivery estimation

Indian e-commerce depends heavily on pincode-level fulfilment. A WebMCP tool can collect the buyer’s delivery pincode and invoke an ONDC-compatible fulfilment or quote workflow. It may return available logistics options, fees, delivery estimates and restrictions.

Because location is personal data, the tool should request only the information needed for the transaction. It should avoid exposing exact addresses to the model when a pincode or serviceability token is sufficient. Final delivery commitments must come from the relevant commerce or logistics system, not from an AI-generated assumption.

4. Cart and order orchestration

Agents can simplify multi-step commerce by managing a cart across search, selection and checkout. WebMCP tools might support adding an item, changing quantity, revalidating price and stock, and obtaining a final quote.

Before order creation, the system should perform a fresh validation because product availability, price and delivery conditions may have changed. The user should see a confirmation summary containing:

• Products and quantities
• Seller or sellers
• Item price and applicable charges
• Delivery fee and estimated timing
• Discounts and coupon conditions
• Taxes where provided
• Payment method
• Cancellation and return terms
• Final amount payable

Only after explicit confirmation should an agent invoke an order-creating action mapped to the appropriate ONDC flow.

5. Payments without exposing sensitive credentials

WebMCP must never require an AI model to handle raw card numbers, UPI PINs, one-time passwords or other payment secrets. A safer design redirects or hands off to a payment interface controlled by the payment provider or authorised application.

The agent may request a payment session or present available methods, but the actual authentication should occur in a trusted user interface. The backend should use payment tokens, signed callbacks and idempotency keys. Order state must be reconciled against confirmed payment and ONDC responses rather than inferred from an agent message.

6. Order tracking, cancellation and returns

After purchase, WebMCP can expose read-only tools for order status, shipment events and expected delivery. It can also guide users through cancellation or returns, but high-impact actions require confirmation and policy checks.

For example, an agent should not cancel an order solely because a user says, “I think I no longer need it,” if cancellation may incur a fee or be irreversible. The tool should return eligibility, consequences and the exact action to be taken, then request confirmation.

Recommended Technical Architecture

A production-grade integration can be divided into six components:

WebMCP tool registry

Define tools using strict input and output schemas. Each tool should have a clear description, risk classification, authentication requirement and supported commerce states. Separate read tools from write tools.

Agent gateway

The gateway authenticates the user and agent session, applies policy, validates parameters, limits tool access and records every invocation. It should reject ambiguous requests rather than silently guessing.

ONDC adapter

The adapter translates internal commerce objects into the relevant ONDC message structures and maps network responses into stable internal representations. Keep protocol-specific logic here so that the WebMCP interface remains consistent when ONDC versions or domains change.

Commerce state machine

Do not model checkout as a collection of independent chat actions. Use explicit states such as search, quote received, cart created, payment pending, order confirmed, cancelled and returned. Each transition should be validated and idempotent.

Identity and consent service

Connect the user, buyer application, seller account and agent session with appropriate consent. Delegated access should be time-limited and scoped. A shopping agent may be allowed to search and compare but not place an order without a new approval.

Observability and reconciliation

Log correlation IDs across the agent, WebMCP gateway, ONDC adapter, payment provider and logistics systems. Monitor failed callbacks, duplicate requests, stale quotes, mismatched totals and delayed status updates.

Security, Privacy and Compliance Considerations in India

The combination of AI agents, commerce data and payments creates a large attack surface. Key controls include:

• Strong authentication: Use secure sessions, OAuth-style delegation where appropriate, signed requests and participant credentials managed in a secrets vault.
• Least privilege: Expose only the tools required for the current task. Do not give a shopping agent unrestricted account or seller administration access.
• Confirmation for high-impact actions: Require user confirmation before order placement, payment, cancellation, returns, address changes or account modifications.
• Prompt-injection defence: Treat product descriptions, seller messages and web content as untrusted data. They must not override system policies or tool permissions.
• Input and output validation: Validate every parameter server-side and filter responses before presenting them to the model or user.
• Personal-data minimisation: Limit use of addresses, phone numbers and order histories to the transaction purpose. Apply appropriate retention and deletion policies under India’s Digital Personal Data Protection framework and other applicable requirements.
• Auditability: Store who authorised an action, which tool was called, what data was used and which network response was received.
• Fraud detection: Add velocity limits, unusual-order detection, device signals and manual review for suspicious activity.
• Price integrity: Recalculate totals from trusted backend data immediately before payment and order creation.

Organisations should also assess consumer protection, tax, payment, grievance redressal and sector-specific obligations. ONDC participation requirements and domain rules must be reviewed alongside general technology controls.

Implementation Roadmap for Indian Startups

A practical rollout can follow these stages:

1. Choose one narrow journey: Start with product search, catalogue assistance or order tracking rather than autonomous checkout.
2. Confirm ONDC role and domain: Establish whether the business operates as a buyer network participant, seller network participant, technology service provider or another permitted role.
3. Build a canonical data model: Define products, offers, locations, carts, orders, payments and fulfilment states independently of the language model.
4. Create read-only WebMCP tools: Validate discovery and tracking before enabling transactional actions.
5. Add policy and confirmation gates: Introduce cart creation, quote refresh and order placement with explicit user approval.
6. Test protocol and failure cases: Cover timeouts, duplicate callbacks, unavailable sellers, price changes, payment failures and partial fulfilment.
7. Pilot with limited users: Measure task completion, incorrect tool calls, conversion, support tickets and safety incidents.
8. Expand carefully: Add returns, multi-seller baskets, merchant automation and personalised recommendations only after governance controls mature.

Common Mistakes to Avoid

• Treating WebMCP as a replacement for ONDC participant integration
• Allowing the model to generate raw protocol messages without validation
• Using browser automation where a stable backend API is available
• Hiding delivery charges, seller identity or return restrictions
• Making purchases from conversational intent without a confirmation step
• Passing full addresses or payment details into model context
• Failing to handle stale quotes and asynchronous ONDC updates
• Building a tool catalogue so broad that the agent can take unintended actions
• Ignoring versioning, certification and operational responsibilities in the ONDC ecosystem

Business Benefits and Limitations

For Indian e-commerce companies, this architecture can reduce friction in product discovery, improve accessibility for users who prefer natural language, and help merchants operate across interoperable channels. Smaller sellers may gain agent-assisted catalogue management and customer support without building sophisticated conversational systems from scratch.

However, WebMCP does not solve supply quality, catalogue inconsistency, logistics reliability or network governance. AI-generated recommendations can also be biased, incomplete or commercially misleading if ranking rules are not transparent. The most useful deployments will combine structured protocol data with clear explanations and deterministic backend controls.

FAQ

Can WebMCP directly connect an AI agent to ONDC?

It can provide the agent-facing tool layer, but a secure integration should use a backend gateway and ONDC adapter. The model should not directly access unrestricted ONDC credentials or send unvalidated protocol messages.

Is WebMCP the same as an ONDC API?

No. ONDC defines interoperable commerce interactions between network participants. WebMCP exposes selected application capabilities to AI agents. It can translate agent requests into ONDC workflows but does not replace ONDC requirements.

Can an AI agent place an ONDC order automatically?

Technically, a system can automate parts of checkout, but production systems should require explicit, informed user confirmation before payment or order creation. Permissions, payment security, audit logs and applicable regulations must also be addressed.

What should startups build first?

Start with read-only product search, offer comparison, serviceability checks or order tracking. These use cases demonstrate value while reducing the risks associated with autonomous payments and irreversible actions.

How can WebMCP protect customer data?

Use data minimisation, scoped permissions, short-lived sessions, server-side validation and payment-provider handoffs. Keep sensitive credentials and unnecessary personal data outside model context wherever possible.

Apply for AI Grants India

Building a secure WebMCP and ONDC integration can be a strong foundation for an AI commerce startup in India. Apply to AI Grants India to explore support and opportunities for developing your AI venture.