The contract never leaves,only the sentence you asked about
Analysts and product owners are handed documents they are not allowed to upload: a signed vendor contract, an HR policy, a client's requirement pack. Khint reads those files on your own machine and sends only the text you act on. No on-device model, no vague promises: here is the line-by-line accounting of what stays and what moves.
“Do not upload this”
is not the same as “do it by hand”
A 40-page master services agreement lands in your inbox and you need the obligations out of it by Thursday. The rule at most companies is not subtle: that file does not go into a browser chatbot, and it certainly does not go into a document-analysis site that keeps a copy. So the work becomes a manual read, and the deadline absorbs it. What people search for at that point is an offline AI, and the honest version of that search has two halves. There is no small model on your laptop that will read a contract as well as a hosted one. But there is no reason the file has to travel for a hosted model to help you, and that is the part Khint takes seriously.
A confidential file, end to end
Five steps between the document on your disk and the answer in your ticket.
Your machine opens the file, not a server
Point an agentat the contract and Khint reads it locally. One extractor covers PDF, Word, PowerPoint, Excel, OpenDocument, EPUB, RTF, and plain text or code, so the document never has to be uploaded for something to understand it. It also checks the file's actual signature rather than trusting the extension: a PDF someone saved as
.txt, or an export with no extension at all, still reads. Containers are read whole up to 25 MB. Nothing about this step touches the network.Only the text crosses the line, and short files cross whole
Once the document is text, the size decides what happens next. Anything up to roughly 10 KB is handed to the agent as-is, so a two-page policy note reaches the model exactly as written. Above that, Khint condenses the document once into a working summary and sends that instead. Either way the payload is text you can reason about, never a file, and never your whole folder: one document per run, the one you picked.
The condensed version is cached on your own disk
That summary is stored in the local database next to your sessions, keyed by a hash of the file's contents. Run the same agent on the same 80-page specification tomorrow and there is no second call: the cached summary is reused. Edit the document and the hash changes, so Khint regenerates. The cache is local only and is never synced to Khint's servers, which means the condensed content of your confidential files exists in exactly one place, your machine. See how document input works for the run-time picker and pinned-file options.
Strip the identifiers that do not need to travel
A signed contract or an HR file is usually sensitive twice over: the commercial substance, and the people named in it. Tick Redact PII on the agent, off by default, and emails, phone numbers, SSNs and IBANs become placeholders before the text is forwarded. You keep the shape of the clause you wanted analysed without shipping the identifiers. More on the redaction switch →
An edit comes back as a new file your machine wrote
For a Word, PowerPoint or Excel file, the model never sees the format at all. Khint hands it a numbered list of text spans, takes back only the spans that changed, and writes every byte of structure itself, so styles, tables, themes and images come out untouched. A PDF is handled in the preview window instead, and the three ways a repaint can degrade (text that will not fit, a box whose background is not one flat colour, a character the embedded font cannot draw) are flagged for you rather than quietly faked. See the in-place editing flow →
What stays, what moves
The same accounting, as a list you can take to your security reviewer.
The part that is not local
The model call goes out. Khint sends the text through its own backend, which forwards it to the model and keeps no copy, and on the Lifetime plan you can point that call at your own provider key from Settings instead, stored in your system keychain. Two other things reach the network by design: reading an image or a scan needs the Vision model, because there is no text to extract locally, and Capture is how you do that deliberately. Everything else in the ledger above stays put, including Memory sessions, which live in a local database and are never synced. If your reviewer wants the whole posture rather than the document path, the privacy-first overview and the security page cover it.
Common questions
Is Khint an offline AI?
Not in the strict sense, and it is worth being precise because that distinction is usually what people are really asking about. Khint does not ship an on-device language model: when an agent runs, the text goes to a hosted model through Khint's own stateless backend, which forwards the request and keeps no copy. What is genuinely local is everything around that call. Your document is opened, parsed and turned into text on your own machine. The condensed version of a long file is cached in a local database. Your action prompts, your Memory sessions and the input and output text of every run stay on the device. An edited document is written locally by the app itself. So the file never travels, only the text you chose to act on does.
Does my PDF or Word file get uploaded anywhere?
No. Khint reads the file on your machine and extracts its text there: PDF, Word, PowerPoint, Excel, OpenDocument, EPUB, RTF, plus plain text and code all go through a local extractor. The bytes of the document itself are never sent anywhere. What can leave is the extracted text, and only for the run you triggered: a short file is sent as the agent's input, and a long one is condensed once into a summary that is sent instead. The one exception is an image file, or a scan with no text layer, because reading pixels needs the Vision model. That is a network call, and Khint tells you so by counting it as a capture.
What happens with a scanned PDF, an encrypted file, or something Khint cannot read?
It says which of those it is instead of failing vaguely. A PDF that yields no text is reported as having no text layer, which means it is a scan rather than a document with selectable text. An encrypted file is reported as encrypted. A file over the 25 MB ceiling is reported as too large, and a permission problem on macOS names the Files and Folders setting you need to grant. Legacy .doc and .ppt are deliberately not parsed: rather than guess at prose it cannot validate, Khint asks you to save the file as .docx or PDF. For a scan you actually need, run it through Capture, which OCRs the image and hands you selectable text.
Where does an edited document end up, and can it damage the original?
The original is never opened for writing. An agent that gives back the edited file writes a new copy into Khint's own outputs folder, named after the source with (edited) appended, and it numbers the name rather than overwriting an earlier result. A preview window shows you the paragraph diff before you keep anything, with Download and Reveal buttons. Khint also refuses jobs it cannot do cleanly instead of half-doing them: an empty document, a single paragraph over 5,000 characters, or a document over 150,000 characters is turned down rather than truncated, because a partly rewritten contract is worse than no rewrite. If the file changes on disk while the model is working, a size and timestamp check catches it and the write is refused.
Read the file without sending the file
Free with 300 credits a month, about 10 AI actions a day. No credit card. Your documents are parsed on your own machine.