Not every question is text. You can attach images, PDFs and ordinary documents — a spreadsheet, a Word file, a slide deck, a CSV, a Markdown or text file, source code, or a zip of them — to a message, or dictate it out loud instead of typing. This is one of the features of Secret Chat AI, the AI aggregator.
Attaching images
Drag a picture onto the chat, or use the attach button. JPEG, PNG, GIF and WebP are accepted, up to 25 MB each. Models that can see will read a screenshot, a photograph of a document, a chart, a diagram or a design and answer questions about it, and in the Image Generator an attached picture becomes the starting point for an edit rather than a question.
Attaching PDFs
PDFs go the same way. The size limit varies by model and is shown in the attach button's tooltip for whichever model you have selected. Useful for the things that are painful to paste: a contract you want summarised clause by clause, a paper you want the method of explained, a manual you want one specific answer out of. A PDF is sent to the provider as a file, because a scanned PDF has to be looked at rather than read, and every model here except DeepSeek reads a PDF natively.
Attaching documents: spreadsheets, Word, PowerPoint, text, code and zip
Everything else is handled differently, and the difference is worth knowing. A .docx, an .xlsx (also .xls and .ods), a .pptx, a .csv or .tsv, a .md or .txt file, a JSON or HTML file, source code in most languages, or a .zip of any of those is converted to plain text in your browser, on your device, before anything is sent. Only that text goes to the model; the original file never leaves your computer. A spreadsheet arrives as one CSV block per sheet, a Word file keeps its headings, lists and tables, a slide deck arrives slide by slide, and an archive becomes one attachment listing each file it contained (and naming the ones it could not read: nested archives, binaries). Images and PDFs inside a zip are not turned into text — they are attached as images and PDFs in their own right, exactly as if you had picked them, on a model that reads them and as far as the twenty-attachment limit allows; any the message has no room for (or the model cannot read, or that are over the size limit) are named in a warning, never dropped silently.
Click the attachment before sending and you see exactly the text the model will receive — no more, no less. The limits: 10 MB per file, 25 MB per zip archive, and about 200,000 characters of extracted text per file (roughly a hundred pages of prose); a bigger spreadsheet is cut at that point and the cut is marked, so neither you nor the model can mistake the first part for the whole. One message can carry up to twenty attachments and about 400,000 characters of document text in total. Legacy binary formats (.doc, .ppt) are not supported — save them as .docx / .pptx first.
A document costs what its text costs as input, and nothing more — there is no per-file fee and no provider-side file storage involved. That is also why documents, like every other attachment, exist on the paid models only: on the free daily quota nothing takes an attachment of any kind (the free modes of ChatGPT and Gemini, DeepSeek USA, Perplexity and the free AI Council pair), because a document can be a hundred pages long and every one of them would be paid for by us.
Which models can read what
Capabilities differ by provider, and the app follows the selected model rather than pretending otherwise: the attach button simply stops offering file types the current model can't read.
- ChatGPT, Claude, Gemini and Grok read images, PDFs and documents — with credits; the free modes of ChatGPT and Gemini take none.
- The free models take no attachments at all. DeepSeek USA and Perplexity show no attach button: a file — a document above all, but an image or a PDF too — is the one thing that can make a single message large, and on a model that is free to use every byte of it is paid for by us. (DeepSeek is also text-only on input, so it could not read an image or a PDF in any case.) Switch to a paid model for anything with a file in it.
- An AI Council takes all three on the branded panels, and nothing on the free pair. A council sends your message to several models at once, so an attachment has to be readable by every one of them: images, PDFs and documents are accepted on every branded panel (Grok reads a PDF through xAI's file upload, which the gateway deletes again the moment the answer is back), and the free DeepSeek + GPT pair takes none of them, for the same reason the free single models don't.
- The Image Generator takes images and PDFs (an attached picture becomes the starting point for an edit) but no documents: a spreadsheet is not an image prompt.
If you attach a file and then switch to a model that can't read it, the app tells you rather than quietly sending a question with the attachment missing.
Voice input
Press the microphone and speak; recording shows a timer, and when you stop, the audio is transcribed and the text is dropped into the message box. You can then edit it before sending; it fills in the input, it doesn't send on your behalf. Convenient for long messages, awkward spelling, and anything you would rather say than type.
Two things to know: voice transcription is a paid feature, not part of the free daily quota, and it needs a browser with microphone recording support, the button tells you if yours can't do it.
What leaves your device
Attachments are stored in your own browser (OPFS) along with the rest of your history, and they're included when you export your data. For a document, what is stored is the extracted text, not the original file. But to be answered about, a file has to be read by the model, so its contents do go to the provider with your message — anonymously, under our gateway's credentials rather than your identity, but in full: an image or a PDF as the file itself, a document as its text. Voice is the same: the recording is transcribed by a provider before it becomes text.
This is where the standing limit bites hardest, because files carry far more than you consciously put in a sentence. A photographed letter has a letterhead and an address on it. A PDF contract has the parties' names throughout. A spreadsheet has the whole customer column, not the one row you had in mind. A scanned report may have a patient identifier in the header you weren't even looking at. Secret Chat AI removes you from your queries; it doesn't remove the data from your messages, and it can't redact a document for you. The app runs its personal-data check over an attached document's text as well as over what you typed, and warns before sending, but that check catches patterns, not meaning. Check what is actually in the file before you attach it, and crop, redact or delete a column when it doesn't need to be there.