In the past few months, Deepinder Goyal has made it abundantly clear that his entrepreneurial ambitions lie in building future-facing, tech-first companies.
His recent $25 million personal commitment to Continue — a research initiative supporting global scientists working on human ageing and longevity — signals a serious investment in health science.
Additionally, he has launched LAT Aerospace to establish a propulsion research team in Bengaluru, aiming to develop indigenous gas-turbine engines for India.
And this vision also shows up with the food delivery platform Zomato, his first major entrepreneurial success and the company where everything began for Goyal.
The team has consistently advanced in AI, whether it’s Nugget, the AI-native customer support platform that handles millions of interactions each month, or Espresso, the high-performance, open-source PDF-generation and signing tool developed in-house.
But last week saw a quieter release on Zomato’s GitHub page: an official MCP server.
MCP is the open-source framework built by Anthropic that the ecosystem has been gung-ho about in the last few months.
In practical terms, MCP connects AI systems or chatbots to real, live data sources and services, not merely for reading, but also for interacting with them.
Instead of relying on whatever the model already knows from training, MCP lets it access fresh, permissioned information and trigger specific actions via an approved interface.
And How Does this Work for Zomato?
After you set up the MCP server, you can use natural language to perform a variety of actions that would typically require manual input on the interface.
This includes fetching specific restaurants based on your instructions, browsing menus, creating carts, ordering food, and even completing payments through QR code integrations.
According to Zomato’s documentation, the MCP server can be accessed via ChatGPT using OAuth authentication, or via Claude — either in the desktop app or the VS Code extension.
So it isn’t exactly a straightforward way to access this feature, unlike some of Zomato’s other features available on a front-end user-facing interface, but rather requires users to perform a manual setup.
Shubham Palriwala, the founder and CEO at Agnost AI, shared a detailed explainer on how one can set up the MCP server.
“With Claude, you just talk. ‘I’m hungry, get me something spicy under 300 rupees near MG Road with good ratings.’ Claude handles the searching, filtering, and discount hunting. One conversation instead of a dozen taps,” read the blogpost.
Palriwala, while testing the MCP server, shared use cases such as locating food of a particular cuisine within a set budget via a single prompt, which then provides options with ratings and delivery times.
Other use cases, such as tracking orders, adding items to the cart, or comparing menus with prompts, are also feasible.
“You can even be vague: ‘I’m in the mood for something spicy’ – Claude will narrow it down with follow-up questions if needed,” added the explainer.
In the final payment step, the Agnost AI report noted that the MCP generates QR codes that must be scanned to complete the payment. It does not automate the payment process because of security risks.
Similarly, over the past few months, other Indian startups such as Zerodha and Razorpay have also incorporated MCP capabilities.
For example, Zerodha’s MCP server allows users to ask the associated chatbot about their portfolio, review statistics, set alerts for when their overall portfolio deviates from particular targets, and more.
Zerodha emphasised that security is a priority with Kite MCP. The AI assistant only accesses data explicitly authorised by the user.
The user’s credentials do not pass through the AI assistant; authentication occurs externally via Kite. Additionally, all actions are read-only, and placing orders is not permitted.
Razorpay’s official MCP server allows users to connect to AI systems and perform payment-related tasks using natural language prompts.
Razorpay founder Shashank Kumar explained in a post on X that the MCP server facilitates swift payment processing, stating, “For example: Type ‘Send ₹500 payment link to Neha,’ and a payment link is generated, shared on WhatsApp, and completed, all within seconds.”
Beyond its core capabilities, MCP opens the door to features that would otherwise remain on the cutting board—ideas shelved for fear of overwhelming users with cluttered interfaces.
Himanshu Upreti, founder of Ai Pallete, sees this as a potential revival of the ‘super app’ concept that has flourished in China and Indonesia through platforms like Gojek.
“But it obviously did not fly [across the globe] because there was still a lot of resistance in terms of going through multiple clicks and selecting all the options—but using MCP, that journey becomes very easy,” he told AIM.
But this will involve developers overcoming several challenges before it becomes a reality. In a conversation with AIM, Parliwala said, “If you feed your LLM with a lot of tool calls, which are multiple MCP servers, it takes up a lot of tokens and might hallucinate.”
Developers will need to find the right balance between giving the LLM full dynamic tool discovery, where it can attempt many possible actions, and enforcing narrower, well-defined workflows that are guaranteed to execute reliably.
Why MCP?
The core feature of MCP, which enables chatbots to access external data, has driven its popularity. However, it offers more than simply retrieving data from outside sources.
In an earlier interaction with AIM, Kirk Kaiser, a software developer and the author of Make Art With Python, said, “It [MCP] is really more about taking actions, and getting the context back from that action.”
This is also the main value proposition presented by the MCP framework, unlike other frameworks such as RAG or GraphQL, which are primarily focused on fetching information.
This has led to numerous use cases for MCP, with many online lists highlighting MCP servers capable of autonomously executing actions across various apps and products. Even organisations like OpenAI, Cloudflare, IBM, Google, and Microsoft have released their own official MCP implementations.
Many developers have also created tools and features with MCP that are not available in several AI chatbots.
For instance, Kaiser developed a tool that lets him enter natural-language prompts and edit a video within Claude. It can also be used to summarise videos and search for specific mentions in the content.
Anthropic announced a few months ago a directory of tools and apps that can connect to Claude via MCP. This allows users to take actions directly within Claude — without manually setting up an MCP server or writing any code.
Similarly, OpenAI recently released the Apps SDK, which enables app developers to embed their app’s functionality directly into ChatGPT conversations, allowing users to invoke third-party services without leaving the chat interface.
The post Your Zomato Experience Will Never Be the Same Again appeared first on Analytics India Magazine.
