Indian quick commerce startup Zepto has developed and open sourced an internal artificial intelligence tool that allows users to place Zepto Café orders using natural language prompts powered by large language models such as Anthropic’s Claude. Built by a Zepto engineer and shared publicly on GitHub, the project highlights the company’s growing experimentation with AI driven workflows. Co founder and chief executive Aadit Palicha amplified the initiative on LinkedIn, pointing to it as an example of the hands on innovation taking place inside the organisation.
At the core of the system is Anthropic’s Model Context Protocol (MCP), which acts as a coordination layer between user instructions and application actions. Instead of functioning like a conventional chatbot that only generates text responses, the MCP based setup can interpret user intent, decide which action needs to be executed, and then carry out that task through a controlled interface. This allows the system to work with live, permissioned information rather than static or simulated data.
A key technical element of the tool is its use of Playwright, a browser automation framework. Playwright enables the AI agent to operate a real web browser, selecting delivery addresses, adding items to the cart, and completing orders in the same way a human user would. By relying on browser level automation, the system avoids deep backend integrations while still demonstrating how AI agents can interact with existing consumer interfaces.
In a demo video shared alongside the project, Pranav Chandra Prodduturi, senior category manager at Zepto, demonstrated Claude placing a dessert order through the MCP server. Reacting to the demo, Palicha wrote, “Interesting projects getting built in our office these days,” underscoring the experimental and exploratory nature of the work.
Zepto clarified that the tool is not integrated into its consumer facing app and is currently intended for internal experimentation and developer exploration. Using the setup requires manual configuration, and security is managed through local browser sessions, environment variables, and existing web level controls. By open sourcing the project, Zepto is offering developers a real world example of how large language models can be combined with automation frameworks to execute end to end workflows.
The initiative reflects a broader trend of companies exploring agentic AI systems that go beyond conversation and move toward action. While still early and non production, Zepto’s MCP based tool offers a glimpse into how natural language interfaces could one day reshape how users interact with commerce platforms.
