PDFCellsA PDF to Excel converter that doesn’t upload your files
PDFCells opens and converts PDFs in your browser: the PDF text, OCR, the checks, editing and every export run on your device, with no upload and no account. Data leaves the browser only when you choose an AI feature, and then only what that feature needs.
What runs on your device
- Opening the PDF or image, including password-protected PDFs and iPhone HEIC photos.
- The pre-check of every page.
- Reading the PDF text (pdf.js) and both OCR engines (tesseract.js and PP-OCR, served from pdfcells.com itself).
- The checks, the editing grid, payee cleanup, your category rules and the built-in categories.
- Bank reconciliation, including reading your ledger file.
- Every export: Excel, CSV, OFX, QBO, Xero CSV.
What is sent, and only when you ask
| You choose | What is sent | Where |
|---|---|---|
| Read a page with AI | That one page as an image, plus the page’s text layer | Our server, which forwards it to Anthropic’s Claude API |
| Fix with AI | The page image and text, the current tables and your instruction | Same |
| “Cheaper, results within a few hours” for a batch’s AI pages | The page images you picked, up to 25 per request | Same, through Anthropic’s Message Batches API |
| Categorise with AI | Cleaned payee names your rules couldn’t place; no amounts, dates or descriptions | Same |
| Name tables | Each untitled table’s header and up to 3 sample rows; no image | Same |
If you mark areas on a page, the image is cropped to the rectangle around your “only” areas (plus a small margin), “ignore” areas are painted white, the text sent is limited to words inside your areas, and a page whose areas match nothing is not sent at all. With two areas far apart, whatever lies between them is inside that rectangle, so mark one area per AI read if that matters. Our server does not store pages, text or results. Anthropic’s commercial terms don’t allow training on API data; Anthropic may keep API data for a limited period for safety monitoring. Details are in the privacy policy.
Check it yourself
You don’t have to take this page’s word for it. Your browser can show every request the page makes:
- Open pdfcells.com, then your browser’s developer tools (F12, or Cmd+Option+I on a Mac), and select the Network tab. Tick “Preserve log”.
- Open a PDF, read it (PDF text or OCR), and download the Excel file. Don’t choose AI.
- Look at the list. The requests fetch things: the app’s own code from pdfcells.com (
/assets/…, including the PDF reader’s worker), the OCR engine and its English model (/tesseract/…), the experimental OCR models (/ppocr/…, only if a page looks weak),/api/config(settings; carries nothing from your file), fonts from fonts.googleapis.com and fonts.gstatic.com, and the sign-in service at clerk.pdfcells.com (plus/api/me, your balance, when you are signed in). If you are signed in, the sign-in service also sends small requests, some of themPOSTs, to keep your session alive; they carry your session, not your file. The Excel file is built inside the page, so the download is not a network request at all. - What you won’t see is any request carrying your file or its contents. That only happens when you choose an AI feature: then you’ll see
POST /api/extract(or/api/categorize,/api/name-tables,/api/batches) with the page image or text in its payload.
The page’s Content-Security-Policy response header (visible on the first request in that list) backs this up: its connect-src lets the page’s scripts open network connections (fetch, XHR) only to pdfcells.com itself and the sign-in service’s hosts (Clerk, its bot protection and telemetry), and the browser blocks any other. Other directives allow loading Google Fonts, Clerk profile images and Cloudflare’s sign-in check. And the code is public: every call that sends data to our server goes through one function, request in src/lib/api.ts, and the policy is in vercel.json. The whole app is open source on GitHub under the MIT licence.
Is it safe to upload bank statements to an online converter?
It can be, but you are trusting the service with account numbers, balances and every transaction. Before uploading client statements, check four things in the service’s own policy:
- How long files are kept after conversion, and whether you can delete them yourself.
- Whether documents are used to train models.
- Where data is stored and which subprocessors see it.
- Whether your engagement letter or firm policy allows third-party processing of client data.
Published policies differ a lot, as of October 2026: Smallpdf says files are deleted from its servers after one hour; BankConv deletes after 24 hours by default; DocuClipper keeps files while you are subscribed and deletes them 90 days after cancellation; Docparser lets you set retention from zero to 180 days. (Sources on the converter roundup.)
Other converters that don’t upload
PDFCells is not the only local option. Tabula is free desktop software for text PDFs. ProperConvert and MoneyThumb’s desktop converter run on your computer and target QuickBooks and Quicken formats. StatementTidy and HarborConvert also say they convert in the browser. Pick by the job: see the comparisons.
Questions
Does PDFCells upload my PDF?
No. The PDF is opened and converted in your browser. Only pages or data you choose to send to an AI feature leave your device.
Do I need an account?
No. Reading PDF text, OCR, checking, editing and exporting work without an account. Signing in is only for free AI pages and credits.
Is it safe for client bank statements?
Without AI, the statement never leaves your computer, so no third party sees it. If you use AI on a page, that page goes to our server and Anthropic’s API and is not stored by us.
Can I use my own Anthropic API key?
Yes. The key is kept in your browser’s local storage and sent with each AI request; our server uses it for that request and doesn’t store or log it.