Two Roads to the Same Shopper: What Happens When Perplexity's Agentic Checkout Meets an ONDC-Only MSME

Three protocols, one closed loop

Start with what shipped in the West over the last five months of 2025 and the first weeks of 2026, because the pace is the point. On September 16, 2025, Google announced the Agent Payments Protocol, AP2, an open specification built on top of its Agent2Agent (A2A) protocol and the Model Context Protocol (MCP). AP2 solves a specific problem: how does a merchant know an AI agent actually has a human's permission to spend their money? The answer is a "mandate," a cryptographically signed digital contract that proves authorization, authenticity, and accountability before a payment clears. More than 60 companies signed on at launch, including Mastercard, PayPal, American Express, Coinbase, and, for India, JusPay as a UPI-focused partner.

Thirteen days later, on September 29, 2025, Stripe and OpenAI announced the Agentic Commerce Protocol, ACP, a "merchant-friendly open standard" that powers Instant Checkout inside ChatGPT. Etsy's US sellers went live immediately; Stripe said Shopify's network of more than a million merchants, including brands such as Glossier and Spanx, would follow. The announcement was explicit that this launched for "ChatGPT users in the United States" buying from "US-based" sellers, with wider geographies promised later, not delivered at launch.

Then, on January 11, 2026, Google added a third layer: the Universal Commerce Protocol, UCP, co-developed with Shopify, Etsy, Wayfair, Target, and Walmart, and endorsed by more than 20 additional companies including Flipkart, Mastercard, Visa, Stripe, and American Express. UCP lets a business show up inside Google's AI Mode and the Gemini app for discovery and checkout. Flipkart's inclusion matters for this article's question, because Flipkart got its route into agentic commerce directly from Google, as a platform-scale endorser with its own backend and its own Merchant Center relationship. Neither ONDC nor the Beckn Protocol appears anywhere in Google's UCP announcement.

The other model: a network nobody has to ask permission to join

India's alternative to platform-by-platform integration predates all three of those launches by years. The Beckn Protocol is an open-source, decentralized specification, originally incubated by the Foundation for Interoperability in Digital Economy (FIDE, co-founded by Nandan Nilekani, Sujith Nair, and Pramod Varma), for discovering and transacting with any compatible service across any domain, retail, mobility, food delivery, healthcare, without a single controlling platform. The Open Network for Digital Commerce, ONDC, is the Indian government-backed network built on top of it: ONDC's own GitHub repository describes the arrangement precisely, with the Beckn Protocol as the "base layer" and ONDC's own specifications as a "network extension layer" adopted to enable transactions across the network.

The scale is real. ONDC's own published network figures show 616-plus cities live, 26 active domains from food and beverage to financial services and agriculture, 306 live network participants, more than 7.64 lakh (764,000-plus) registered sellers and service providers, and more than 16 million cumulative orders as of May 2025. The government is also directly subsidizing MSME onboarding: the MSME TEAM scheme, run through the Ministry of MSME's RAMP program with a budget of Rs. 277.35 crore over FY2024-25 to FY2026-27, targets bringing five lakh micro and small enterprises onto ONDC, half of them women-owned, funding catalog preparation, account management, and logistics support.

The design principle that makes this work is also the one that creates the gap this article is about. An MSME on ONDC does not build or operate its own storefront, checkout, or product feed. It works through a Seller Network Participant, an app that speaks Beckn's protocol on the seller's behalf. That seller's entire digital commercial identity lives inside a network schema built for network-native buyer apps, not inside a Shopify store, a BigCommerce account, or a Stripe-connected checkout, the exact kinds of merchant backends that AP2, ACP, and UCP were each built to plug into.

Where a bridge exists, and where it does not

The clearest sign of institutional appetite for connecting Beckn networks to AI agents came on November 3, 2025, when Beckn Labs and Google Cloud announced a partnership around Beckn-ONIX, an open-source "universal adapter" for Beckn-based networks. Google Cloud's own press release states that Beckn-ONIX cuts network onboarding time from four to five months down to roughly 45 minutes, and that developers can "integrate AI-native capabilities into Open Networks from day one" using Google's Agent Development Kit and Gemini models. The release cites a live example: an Open Network Stack for Agriculture in Uttar Pradesh, built on Beckn-ONIX, that has engaged more than 12,800 farmers.

That is a genuine and useful development, and it is not the same thing as the bridge this article's title describes. Beckn-ONIX and the Agent Development Kit let a developer build an AI agent that operates inside a Beckn network, a Gemini-powered assistant, say, that speaks Beckn to a Beckn Provider Platform. It does not make that same ONDC seller discoverable to an outside agent that speaks AP2, ACP, or UCP, the protocols actually wired into ChatGPT, Perplexity's Comet browser, and Google's own AI Mode and Gemini app for general consumer shopping. Those are different conversations happening in different protocol languages, and nothing announced so far translates between them at the network level.

The closest attempt at that specific translation is a community project, not an official one. On March 29, 2025, a developer using the handle RakeshVT posted a GitHub Discussion in the Model Context Protocol organization titled "Introducing Beckn-MCP Integration: Open Commerce for AI Agents," describing a reference implementation that lets any MCP-compatible AI assistant discover and transact with Beckn networks, explicitly framed as a way to keep AI shopping from consolidating inside a handful of proprietary "walled gardens." The author describes it as "functional" but not thoroughly reviewed, and the thread shows interest from FIDE's technical team, not an adoption commitment. It is a proof of concept sitting in a public discussion thread, not shipped infrastructure any ONDC seller can rely on.

What a centralized platform can do that an open network cannot, yet

The contrast sharpens against what India's large, centralized platforms have already done on their own. Swiggy's own developer documentation describes three live MCP servers, covering Swiggy Food, Instamart, and Dineout, offering 35 tools total, letting AI assistants such as ChatGPT, Claude, and Gemini search, build a cart, and place a food or grocery order or a restaurant reservation through natural-language prompts. Swiggy could do this because Swiggy controls one backend and wrote one connector for it. Flipkart's UCP endorsement runs the same way: one company, one backend, one direct relationship with Google.

An MSME that trades only through ONDC has no equivalent lever to pull. It cannot write its own MCP server or strike its own protocol partnership, because by design it does not operate the backend a Western agent protocol expects to talk to; the Seller Network Participant app does. The bridge, if one gets built, has to be built once, at the network level, by ONDC, Beckn Labs, or a partner working with them, not by each of the seven-plus lakh sellers individually. That is precisely why nobody should expect this gap to close through the accumulation of individual seller effort, the way a single merchant might improve its own Shopify listing. It closes at the network layer or it does not close.

What would actually close the gap, and what would not

Beckn-ONIX's plugin architecture is the more plausible foundation for a real translation layer than a fresh build would be, since it already normalizes Beckn's search, select, init, and confirm calls into API shapes external systems can consume. But an adapter that talks to AI agents inside a Beckn network is not the same engineering problem as a certified endpoint that lets an AP2 mandate, an ACP product feed, or a UCP business listing resolve to an ONDC seller's live catalog and confirm an order against it. That second problem also is not purely technical. AP2's mandates exist specifically to answer who authorized a purchase and who is liable if something goes wrong; ACP ships with Stripe as a processor and a defined merchant-of-record structure; UCP is anchored to Google's existing Merchant Center relationships. Beckn's own settlement and liability model was not built around any of those three assumptions, so a real bridge has to resolve that mismatch, not just the message format.

None of this means the gap is permanent. The Google Cloud and Beckn Labs partnership shows real institutional interest in AI-native Beckn infrastructure, and the FIDE team's engagement with the independent Beckn-MCP project shows the idea has reached the people who could formalize it. As of this writing, no announced, shipped, officially maintained bridge yet lets an AP2-, ACP-, or UCP-speaking shopping agent discover and check out from an ONDC-only seller. Anyone claiming otherwise should be asked for the specific announcement and the date on it.

What this looks like from inside the gap, today

For the roughly 7.86 crore MSMEs registered on India's Udyam portal as of February 2026, a sector that the Ministry of MSME's own data credits with 31.1 percent of GDP and 48.58 percent of exports, the immediate stakes are not abstract. Generative-AI referral traffic to US retail sites rose 1,200 percent between July 2024 and February 2025 according to Adobe Analytics, and trade-press coverage of Perplexity's own reporting puts its shopping-query volume at fivefold growth since its shopping feature launched. That growth is arriving on rails, AP2, ACP, UCP, that an ONDC-only seller currently cannot ride.

Two things remain true regardless of when a Beckn-to-agent bridge ships. First, being listed on ONDC still means being discoverable through ONDC's own family of buyer apps, a real and growing channel in its own right. Second, machine legibility on the open web, a consistent business identity, structured information, genuine reviews, content an AI system can verify and cite, is a separate layer from checkout-protocol plumbing, and it is the layer that determines whether Perplexity, ChatGPT, or Google's AI Mode names a business in an answer at all before any agent ever attempts a purchase. That layer is measurable now, independent of which commerce protocol eventually gets built to connect to it.