Read a receipt into fields you can keep
Pick a photograph or a PDF. One model call reads it into typed fields, the app checks the arithmetic, you correct the rest.
The app keeps nothing on the server: your upload lives in memory for the length of the request and reaches no disk and no database. This session's receipts live in this browser, survive a reload and end when you close the tab.
Or scan one of three samples
All three receipts are fictional. The merchants, the addresses and the card digits belong to no real business and no real card. Press one and the app reads it the way it reads a file you pick yourself.
Grocery till receipt (fictional)
Eleven items whose amounts fall 0.90 short of the printed subtotal, so the arithmetic warning appears without hunting for a flawed receipt.
Restaurant bill (fictional)
A meal with a tip and two separate tax lines, where every printed figure adds up.
Scanned invoice (fictional)
A PDF holding a photographed page and no text layer, so the app rasterizes it and reads the words with OCR.
The receipt
Nothing chosen yet. The photograph you pick appears here.
Waiting for a receipt
Pick a receipt file and press extract. The fields appear here.
What happens to your receipt
Six steps, in order. One of them runs on a server. The other five run in your browser, which is why no receipt has to be stored anywhere to be read.
Step 1. You hand it a file
The picker takes JPEG, PNG, WebP and PDF up to 8 MB. A PDF is rasterized in your browser at twice its page size, and a photograph too large for the request body is re-encoded and shrunk until it fits. What the document pane shows you is the image the model reads.
Step 2. One model call reads it
The image goes to gpt-6.1-sol with Structured Outputs in strict mode, and the schema it has to satisfy is the field set itself. Zod validates the answer again on the server. No JSON is parsed out of prose, and nothing retries behind your back.
Step 3. The app finds each value on the page
Nothing asks the model where a value sits, because it would answer with numbers that look precise and are not. The browser reads word boxes off a PDF's text layer, or runs tesseract.js in a worker when the page has none, then scores runs of adjacent words against the characters each field was read from. Below a score of 0.72 it marks nothing rather than marking the wrong line.
Step 4. The arithmetic is checked in code
The line items have to sum to the printed subtotal, and the subtotal plus the taxes plus the tip has to equal the total. Both comparisons are exact decimal equality over scaled integers, with no cent of tolerance, so a receipt that rounds its own tax line says so. A mismatch raises a warning naming the difference on both fields it involves, and the app never rewrites a number to close a gap.
Step 5. You correct what it got wrong
Every value is editable where it sits. Blur or Enter commits an edit, Escape puts back what the field held, and the field's own validator off the schema refuses a date written the wrong way round. A committed edit reruns both checks, clears the confidence on that field and marks the cell as something you typed.
Step 6. You take the data out
The JSON download carries the receipt, the warnings and the list of fields you typed into, with every confidence and source text intact. The CSV carries one row per line item with the receipt's own fields repeated on each row, which is what lets a spreadsheet group or pivot on any of them. Nothing reaches a server, and the receipts this tab holds end when you close it.
What the app reads, and how often it reads it right
The first three figures were measured on 2026-10-07 against gpt-6.1-sol, over 40 labelled fixtures, and the run is written down fixture by fixture in fixtures/REPORT.md. Each tile says what the set was. The other three are read out of the code that enforces them, so a cap that changes changes this row with it.
98.7%
of fields read correctly
1,662 of 1,684 compared fields matched their label across the 40 fixtures, measured on 2026-10-07 against gpt-6.1-sol. The 40 fixtures are drawn and then damaged in code rather than photographed, so synthetic fading is not thermal paper and the set carries no handwriting and no language but English. Twenty-two fields came back wrong and fixtures/REPORT.md names every one of them.
100.0%
of totals read correctly
The app read the printed total on all 40 fixtures, 40 of 40. It is broken out beside the date because those are the two values a person checks first.
85.0%
of dates read correctly
34 of 40, the weakest field in the set. All six misses are one mistake: a printed date whose day is 12 or under read in the other day-month convention, so 05/06 can come back as either the fifth of June or the sixth of May. Nothing on the paper says which convention it used, and the app sends no locale with the image. The eight invoices printing an ISO date read 8 of 8, so check a slash-separated date before you trust it.
12
fields per receipt
The merchant and its address, the date and the time, the currency, the subtotal, the taxes, the tip, the total, the payment method, the card's last four digits and the line items. Each value arrives with the model's confidence in it and the characters it was read from.
4
file formats
JPEG, PNG, WebP and PDF, up to 8 MB. A PDF is rasterized in the browser and the first page is what the model reads.
40
scans a session
The allowance a browser tab holds. It rests on a value the browser asserts, so it is a courtesy rather than a control, and the app says so where a refused visitor reads it.
The app reading a receipt
Both captures are of the running app on one of the sample receipts. Click either one to open it at full size.
A sample receipt read, with the printed total marked on the image and the arithmetic warning naming the 0.90 gap.
Every line item as an editable cell, with the control that adds a row the model missed and the JSON and CSV exports under it.
Try it on a receipt
Three fictional receipts sit under the picker, so you can watch the arithmetic warning appear without hunting for a flawed receipt of your own. Nothing you upload is stored, and no account is asked for.