Swiggy is opening its food, Instamart, Dineout, and Scenes services to external AI agents for discovery, ordering, reservations, and tracking.
The move positions Swiggy to capture transactions from consumers who increasingly turn to AI assistants to decide what to buy and where to eat.
Swiggy has four MCP servers with 66 tools connecting its services to AI agents, alongside systems to switch models and restrict agents’ access.
Swiggy
is stepping up its focus on making its commerce services accessible to external AI agents, enabling users to discover food, build carts, place orders, and book restaurant tables through AI assistants beyond its own app.
Speaking at Inc42’s inaugural ‘The CTO Summit 2026’, Swiggy CTO Madhusudhan Rao outlined how the company is making its services accessible to agents rather than assuming every customer’s shopping journey will begin within its app.
The company has four Model Context Protocol (MCP) servers — connectors that let AI tools interact with its services — with 66 tools spanning food, Instamart, Dineout, and Scenes. These cover discovery, menus, carts, ordering, reservations, and tracking.
The approach allows external AI assistants to handle users’ requests while tapping into Swiggy’s commerce infrastructure to carry them out.
For instance, Rao said users could connect Swiggy’s MCP to an LLM and ask it to help plan meals around a dietary schedule. “Today you can do that by using our MCP with whichever LLM that you use,” he said.
Such connections could help Swiggy cater to individual preferences and requirements that would be difficult to accommodate through fixed features within its app. Instead of building a separate interface for every use case, the company is making its underlying services available for AI agents to use.
“Building out core capabilities which are simple to use, can be invoked by other services, humans, agents, that continues to stay a focus,” Rao said, referring to functions such as ordering and discovery.
Alongside opening its services to external agents, Swiggy is reworking its internal AI systems to reduce dependence on any single model.
Rao said the company had previously built a contact-centre agent around one model. When the provider faced capacity constraints, moving away from it took nearly a month.
That experience prompted Swiggy to invest in systems for evaluating and experimenting with alternatives, making it easier to change models without rebuilding the entire workflow.
“Treat the model as replaceable, invest in the workflow, the evals,” Rao said.
Swiggy now uses an LLM gateway to direct tasks to different models and tests alternatives in live operations. It also uses models across cloud providers, choosing between them based on availability, response time, and cost.
The aim is to retain the systems built around an AI task while allowing the model performing it to change.
Opening commerce services to AI agents also raises questions about what they should be allowed to access and which actions they can take.
Swiggy is deploying agents in separate environments with restricted access, rather than placing them alongside its core production services.
“We are actually deploying them into separate production-like fabrics but with very restricted ingress and egress rules,” Rao said.
These controls limit what can enter or leave the environments in which agents operate. The company is also working through questions around agents’ identities, permissions, and trust as it expands their use.
Internally, Swiggy is using AI to help new delivery partners complete onboarding. Its in-house assistant guides riders through the process in their local language.
“You will actually get a buddy which can walk you through all the steps in your local language,” Rao said.
He added that the assistant is improving rider net promoter scores (NPS) and the onboarding funnel, although he did not share figures.
Source: Inc42
