Every document you send to AI weighs far more than the model needs to read.
Koy Decoder opens the document, strips out everything that is not information and returns compact, structured text. The model gets the same data at a fraction of the tokens. Nothing uploaded, no account, no limits.
Extract the text with OCR
Recognition runs inside your browser, like everything else. The first run downloads about 3 MB of engine and language data, and takes a few seconds per page.
What to include
Uncheck anything the model does not need to read. Tokens update instantly.
Download the result
The recommended format depends on the document you loaded.
There is no text to download until the document goes through OCR.
The document never leaves your computer
The conversion runs inside your browser. There is no server receiving the document.
Close the tab and no copy is left anywhere.
No account or email required, and no limit on documents.
What it does under the hood
One reader per file type
It identifies the format and processes it with the right reader: spreadsheet, PDF with text, scanned PDF, image or Word document.
Strips what is not information
It drops empty cells and columns, filler rows, headers repeated on every page, pagination and layout leftovers. No cell with content is lost.
Keeps tables as tables
Flattening a document into running text breaks column alignment and leads the model to invent relationships. You get aligned tables and named fields, in the cheapest representation that keeps the structure.
Convert once, ask many times
The document is converted once and reused across every later query. The saving multiplies by the number of queries, and response time drops.
Frequently asked questions
Is the document uploaded to a server?
No. The conversion happens inside your browser with code that was already downloaded. The file is never transmitted anywhere, and closing the tab leaves no copy behind.
Why does it save tokens if the content is the same?
Because much of what gets sent today is not content: empty cells, spacer columns, headers repeated on every page and layout leftovers. That takes up tokens and tells the model nothing.
Is the token count exact?
It is an estimate of roughly 4 characters per token. The savings percentage is reliable because it compares two versions of the same text; the exact absolute figure needs the official counter from your model provider.
Which formats does it accept?
Spreadsheets (XLSX, XLSM, CSV), PDFs with text, Word (DOCX), HTML and plain text. Scanned PDFs and images are measured but not yet converted: they need OCR.
What if my file is already clean?
The tool says so and reports no savings. If a document is almost entirely data, there is nothing to take out of it.
Do you keep anything I convert?
No. There is no server receiving the document and no database to store it in, and the page asks for no account or email.