Turn a PDF bank statement into a spreadsheet without uploading your finances to anyone. The statement is read in your browser, the transactions are detected, and you get an editable table to check and correct before exporting to CSV or Excel. That review step is deliberate: bank layouts vary enormously and a statement is a document that happens to contain a table, not a data format, so the output is a proposal rather than an answer. Totals for money in, money out and the net are shown so you can check them against the statement itself before you trust anything. Drop several statements at once and export them all as one ZIP. Your financial data never leaves your device — you can disconnect from the network and this still works.
| Column | Taken from | Notes |
|---|---|---|
| Date | The token that opens the row | Day-first, month-first or ISO, settled across the whole statement |
| Description | Everything between the date and the amounts | Wrapped lines are joined on |
| Amount | The movement column | Minus, parentheses and DR all read as negative |
| Balance | The rightmost number, when there is one | Detected across the statement, not per row |
| Totals | Computed from the rows | Shown so you can check the parse before exporting |
No. The PDF is read by your own browser and never sent to a server — no file, no transaction, no account number. You can turn off your network connection and the tool still works, which is the simplest way to prove it to yourself. This is the whole reason to use a local tool for this rather than one of the services that ask you to upload.
No. A scan is a picture of a statement with no text in it, so there is nothing to extract. The tool tells you when a PDF has no text layer instead of silently producing an empty table. For a scan you would need OCR first.
Because bank layouts differ enormously and no set of heuristics gets all of them right. The review step is the point of the tool, not an inconvenience — the totals are shown so a wrong parse is visible immediately, and every cell can be edited, added or deleted before export.
From the rest of the statement. Any date with a day above the twelfth settles the order for the whole document, because a number above twelve cannot be a month. Only when nothing in the statement decides it does your day-first or month-first setting apply, and the tool tells you which reading it used.
It handles the three ways statements write a negative — a minus sign, parentheses, and a trailing DR — with the marker taking precedence over a bare sign. Reading a parenthesised amount as positive would turn a payment into a deposit, so this is worth checking against the totals shown above the table.
Yes. Drop as many as you like, or a whole folder. Each gets its own reviewable table, and you can export them individually or all together as a ZIP of CSVs.
The layout probably does not match the heuristics. You can add rows by hand, or extract the raw text with the PDF text extractor and tidy it up in the CSV cleaner.
No. The CSV is written with a byte-order mark, which is what tells Excel the file is UTF-8 rather than guessing at a legacy encoding.
pdf.js extracts positioned text, and every heuristic downstream works on plain `{ text, x, y }` records so the parsing can be reasoned about and tested without a browser. Fragments are bucketed into visual lines within a vertical tolerance, because two pieces of the same printed row rarely share an exact y, and then sorted left to right; pages are offset in the y ordering so a statement that runs across a page break stays in sequence. A line becomes a transaction when it opens with a date and carries at least one amount, and lines without a date are treated as continuations, since long payee names wrap and silently dropping the tail leaves descriptions that are wrong rather than merely short. Three details decide whether the output is trustworthy. The first is the date order: 03/04/2026 is the fourth of March or the third of April and nothing in the token itself says which, so guessing per row would quietly scramble a third of a year's transactions. The whole statement is used instead — any first field above twelve cannot be a month, any second field above twelve cannot be either, and a four-digit lead means ISO — and only when the document settles nothing does the stated default apply, because a declared assumption is correctable while a guess looks like a fact. The second is the sign. Statements write a debit three different ways: a minus, parentheses, and a trailing DR. Reading (1,234.56) as a positive number turns a payment into a deposit, which is the worst thing this parser could do without saying so, and all three are handled with DR overriding a bare sign rather than compounding it. The third is telling money from digits that merely look numeric: grouping separators vary in width, since 1,234,567 elsewhere is 12,34,567 in India, so separators are stripped rather than matched to a fixed group size, and a plain run of digits with no decimals and no grouping is rejected past six characters because a card or reference number is far more likely than an amount that large written bare. That same money-shape test is what keeps a wrapped description carrying a reference number from being mistaken for a transaction row. Whether a balance column exists is decided across the statement rather than per line, so one row that happens to print two numbers cannot change how every other row is read. Nothing is exported without a review step: credits, debits and the net are shown so a parse that has gone wrong is visible against the statement's own totals, and every cell is editable before the CSV or XLSX is written. Export reuses the CSV writer the other CSV tools already use, with a byte-order mark so Excel does not mangle non-ASCII payee names.