Uploading statements
Which files work, what the upload screen tells you, and what to do when a format isn't recognized.
Everything in the app comes from files you download from your bank yourself. Nothing is linked, no credentials are brokered to a third party, and no permanent connection to your accounts exists to be abused later.
What you can upload
CSV or PDF, up to 5 MB per file. You can drop a batch at once; each file is processed independently.
Thirteen CSV layouts are recognized with no setup at all, covering Ally Bank, American Express, Capital One, Chase (four variants), PNC, Revolut, TD, and Wells Fargo (checking and credit card).
Two PDF statement types are read directly: Chase monthly credit-card statements and Wells Fargo monthly checking statements. PDFs are parsed by their layout — the Wells Fargo parser reads the horizontal position of each amount to tell a deposit from a withdrawal from a running balance, because on the page they are all just positive numbers in different columns.
Anything else goes through the mapping flow below, once, and is recognized from then on.
The upload screen
Beyond the drop zone, the screen answers "am I up to date?".
Statement freshness groups everything you have ever uploaded by institution and shows the last six months as a coverage strip: which months have a statement, which don't, and how many months a source is behind. A statement covering January through March marks all three months, not just March.
A source whose most recent upload failed is flagged as needing attention rather than counted as fresh — you already tried, and telling you to try again is the wrong nag.
Saved formats lists the CSV layouts this workspace has taught the app to read. Deleting one forgets how to read that format; it does not touch transactions already imported through it. That's the recovery path if you ever confirm a mapping wrongly: delete it, upload the file again, and map it correctly.
When a format isn't recognized
The upload lands on unsupported, no transactions are written, and — for a CSV — the mapping editor opens.
Mapping a CSV
The editor asks which column is the date, which is the description, and which carries the amount (either a single signed column, or a separate debit and credit pair). If the file has no header row, clear the "first row names the columns" toggle and the columns are labeled by position instead. Headerless files are first-class here: three of the built-in formats have no header either.
You never write a regular expression. The rule for recognizing this format next time is derived from the mapping you just made — the column count, plus the shape of the cells where you said the date and amount live.
Where the app can, it proposes a mapping to start from. That proposal is a starting point and never an answer: every field stays editable, the preview underneath is your own file's bytes parsed through the mapping you are currently looking at, and nothing is imported until you confirm. With no proposal at all the same editor works exactly the same way — the manual path is the floor, not a fallback.
The proposal is produced by a language model that is shown the header and at most ten sample rows with long digit runs masked, and can only answer with column numbers. It has no way to return a date, a description, or an amount, so no value you see in the preview came from it.
The check that matters
Before a mapping is saved, and again on every future upload that uses it, the parsed rows are checked for balance continuity: each row's balance, minus the one before it, must equal that row's credit minus its debit.
That check exists because it's the only one that catches the two things a mapping actually gets wrong — debit and credit columns swapped, or the sign convention inverted. "The dates parse" and "the amounts parse" are satisfied just as happily by a backwards mapping.
If continuity fails on an already-saved profile, the file is marked failed with diagnostics and zero rows are written. That is deliberate: under a mapping you already confirmed, a continuity failure is drift worth surfacing, not a guess worth fixing silently.
Dates come from a fixed list of formats rather than a pattern you type, and every format added since the original parsers is strict — it rejects a date that rolled over rather than accepting it. That's what stops a DD/MM file mapped as MM/DD from importing a year of transactions on quietly wrong days.
PDFs and other files
There is no mapping editor for a PDF — reading one is code, not configuration. Instead the screen offers a short form: which bank it came from and what kind of account. That report goes to whoever maintains the parsers, and it is the only thing that turns an unsupported statement into a supported one.
Uploading the same statement twice
Harmless. Every transaction's id is derived from its date, description, and amounts, so a row that is already in the workspace is recognized and skipped. Overlapping statements — a December file and a January file that both contain New Year's Eve — merge cleanly.
That protects against re-importing the same file. It does not catch the same charge arriving twice through different routes, such as a credit-card CSV and the PDF statement covering the same month, where the two descriptions differ. Duplicates is the tool for that.
What happens to the file
The original bytes are archived so a statement can be re-parsed later — for instance after a parser improves, or after you fix a mapping. They're stored under your workspace and only reachable through the app, by a member of that workspace.