Your nutrition app doesn't have an MCP server.

The file is parsed in your browser, not read by the AI — so the rows can't be mistyped on the way in, and you see the exact meals before any of them are written. Exports from MyFitnessPal, Cronometer, Lose It!, and MacroFactor have their columns recognised by name; any other CSV works too, you just point the mapper at each column once. What comes across is the date and time, food, meal, calories, protein, carbs, fat, fiber, total sugar, and caffeine in milligrams — and alcohol as well, if you've switched alcohol tracking on first.

The awkward parts of real export files are handled: DD/MM/YYYY and MM/DD/YYYY dates, energy in kilojoules as well as kilocalories, semicolon-delimited European files whose numbers use comma decimals, quoted fields with line breaks inside them, trailing totals rows, and deleted-row flags. Column headings don't have to be English either — a German export's Kalorien or Ballaststoffe is recognised, and fiber, sugar, and caffeine are matched in Spanish, French, Italian, and Dutch too. Where a file is genuinely ambiguous — 05/06 could be May or June — the importer shows its reading next to a row from your own file and asks you to confirm rather than guessing. And each row carries a content fingerprint, so re-importing the same file reports the meals as already logged instead of duplicating them.