Fava is primarily a user interface for Beancount. An MCP server (for Beancount) is explicitly not a user but a tool interface so I don't think that fits here.

I'm using fava as a front-end to my AI coding agents already.. I've written a bunch of wrappers. It's really neat what fava does, but stretching a slight bit the API surface, would position it excellently for being an AI-backed tool.

The best would probably be a CLI tool, to interface with a fava server.. so no MCP even.. some flags for authentication depending where you host the thing, and then, list files, get the file, grep the file (perhaps respecting the transactions boundaries?), edit an entry, add an entry, how extraordinary would that be.. then exposing the query language, different pages, list of accounts, fetch some metadata goodness in there.. wow wow wow :)

I was thinking, preparing, planning, the mcp for beancount.