Building an MCP server to ask AI about my finances

I got tired of opening three different apps to answer "how much did I spend on Uber this month". Instead of building yet another dashboard, I built an MCP server: instead of a screen, my finances become something any compatible AI assistant can query directly, in conversation.

MCP (Model Context Protocol) is an open protocol that lets an AI client — Claude Desktop, Claude Code, this very conversation — connect to a server and call typed tools on it, with no UI needed on my end. The client reads the list of available tools and decides on its own which one to call to answer the question.

The data source is Pluggy, via Open Finance: it handles the annoying part (connecting to each bank, keeping tokens fresh, handling auth errors) and hands me real, up-to-date data, with the user having consented to the connection. My server runs on Cloudflare Workers — cheap, close to the user, no cold-start pain for responses this size.

The design decision that paid off the most: no generic "ask anything about your finances" tool. I split it into small, specific tools — account summary, transactions, card statements, spending by category, loans, investments, indicators — each returning a predictable data shape. That takes the guesswork out of the AI's hands, and takes the job of writing a giant schema-explaining prompt out of mine.

The result: I can ask "how much did I spend on Uber in October versus September" in any MCP client and get a real answer, pulled from my actual bank account, with no app to open. The same server could eventually power Órbita too, my other financial-planning project — the upside of separating the data logic from the interface is not having to pick just one.