If your users live in Claude or ChatGPT, an MCP server matters more than your website

If the people you built something for spend more of their day inside Claude or ChatGPT than on the open web, the way to reach them may not be a better website — it may be an MCP server that lets those assistants call your product directly. That's the shift Madan Chaukulkar, founder of Koko Finance, made after building an app that helps people actually use the credit cards already in their wallet: data on roughly 150 cards, covering rewards, APRs, and benefits.

Instead of grinding to pull users onto Koko's own web app, he turned Koko's data layer into an MCP server. Practically, that means a question like "which of my cards should I use in London next month" can be answered by Claude directly, calling Koko in the background, without the person ever opening Koko's website.

Two decisions made that work. First, he stopped positioning Koko as a competitor to consumer finance sites and started treating it as infrastructure — a data layer other fintech products could call instead of scraping the same data themselves, which is a different, far less crowded market than fighting for direct traffic. Second, the build itself wasn't a large engineering lift: he directed an AI coding tool the way he'd managed engineering leads in a prior finance career — stating the goal, having it review the existing codebase, and asking what was missing — rather than writing the implementation himself.

The broader takeaway for anyone sitting on an app, a dataset, or a workflow: if the assistant is now the interface your users actually use, building a way for that assistant to call your product directly can matter more than optimizing the page a shrinking share of people will ever visit.