Turn a client's receipts into a spreadsheet — without uploading them
A client sends a folder of phone photos and PDF invoices. Somebody has to turn that into rows: vendor, date, total, tax. The AI tools that do this well want you to upload the paperwork — which is exactly what your engagement letter, and often your client's, says you will not do.
The situation
- The documents are somebody else's financial records, and most engagement letters treat sending them to a third-party service as a disclosure you need permission for.
- The work itself is not hard, it is just long: the cost is typing, not thinking.
- Off-the-shelf expense apps want an account per client, a monthly fee and a copy of the documents on their servers.
How it works
1. Open the agent in your browser
Nothing to install and no account needed to try it. The model downloads once into your browser's own storage.
2. Drop in the PDFs and photos
They are read in the page. A digital PDF invoice is parsed exactly, text and all; a photo is read by an OCR model running on your machine.
3. The numbers come from rules, not from the model
Totals, tax, dates and currency are matched with deterministic patterns in Python, and the agent flags a total it had to guess rather than asserting one. The language model is only used for the vendor name and a spending category.
4. Export the table as CSV
It saves from the page into your own downloads, ready for whatever you actually keep the books in.
The agents that do this
Receipt & Invoice Extractor
Based on the Receipt & Invoice Extractor template.
Private Quote & Material Estimator
Drop competing contractor quotes and compare them side by side — totals, inclusions, exclusions — then ask questions across all of them, fully private and in-browser.
Open one and it runs in this browser — nothing to install, and no account needed to try it. You can also browse everything else people have published.
What never leaves your computer
- The images and PDFs are opened by JavaScript in your tab — there is no upload step, because there is no server to upload to.
- Our servers see the request for the page itself and nothing else: not the documents, not the extracted table, not the questions you ask about it.
- You can download the agent as a single HTML file and run it with the network off entirely, once the model is cached.
Why this is structural, not a policy
There is no server-side AI on AgentOp: the model runs in your browser tab, so your documents have nowhere to be sent. The full account of what our servers can and cannot see is on the trust page.
What it will not do
- Digital PDF invoices are extracted exactly. Photographed receipts are read by a small on-device OCR model: vendor names and item lines usually survive, and the totals block sometimes does not — when it cannot read a total the agent says so instead of inventing one.
- The OCR model reads English printed text. A handwritten receipt, or one in a non-Latin script, will not come out usefully.
- It is an extraction tool, not an accountant: it does not classify for tax, reconcile against a ledger, or file anything.
Will it run on my machine?
The on-device option needs a recent computer with a graphics card, or an Apple Silicon Mac, and Chrome or Edge — phones and iPads cannot run these models at all. Rather than guess, the browser check measures your actual machine in about a second and tells you which models it can run, before anything downloads. If the answer is no, the same agents work with your own OpenAI or Anthropic key instead — at the cost of the text leaving your machine, which for confidential material is the trade-off this page exists to avoid.
Questions
No. The files are read by code running in your browser tab. AgentOp's servers never receive the documents, the extracted table, or the questions you ask about them — there is no server-side processing step to send them to.
No. The on-device option is free and needs no key: an open model downloads once into your browser and runs on your own hardware. If you would rather use OpenAI or Anthropic, you can supply your own key and the browser calls them directly, but then the text does leave your machine — for confidential client paperwork the on-device option is the point.
Yes, on the on-device option. Download the agent as a single HTML file; once the model has been fetched once and cached, the file runs with no network connection at all.
It reports what it managed to extract and marks the total as a guess when it had to infer one, rather than filling the gap with a plausible number. Amounts are matched by explicit rules — a bare integer is refused as money, because receipts are full of barcodes and product codes that look like totals.
The on-device option needs a reasonably recent computer with a graphics card, or an Apple Silicon Mac, and Chrome or Edge. The browser check page tests your actual machine in about a second and tells you before anything downloads. Phones and iPads cannot run the on-device models at all.
Related
Other jobs like this
The same idea applied to other kinds of confidential paperwork.
Will it run here?
A one-second test of your machine, before any model downloads.
Trust & security
Exactly what our servers see, and what they never receive.
Build your own
Change the wording, the fields or the rules — it is a template.