Documents

Hand it the file,get the file back

PDF, Word, PowerPoint, Excel and the rest, read straight off your Mac. Pick the file when you run or pin one for good, and choose what comes back: the text, or the same document rewritten with its design intact.

Drop the file on the palette
and the work happens on it

An agent whose input is a document reads the file instead of your selection: no copying, no pasting a spec into a chat window, no window to keep open. Pick it from the palette, or drop it there, and the agent runs on what the document says.

KhintPalette
Sprint 24
Documents

The file goes in, the file comes back

The picker is free and opens before anything runs, so a canceled pick costs nothing. Here the agent is set to hand back the document itself: the paragraphs turn over, the styles and images stay exactly where they were, and what you get is a new file with Download and Reveal on it.

Whatever landed
in your inbox this morning

The formats you actually get sent, read locally rather than uploaded somewhere. Khint looks at the bytes rather than the file name, decides once per agent where the file comes from, and lets a whole workflow start from one.

Every format you actually get sent

PDF, Word, PowerPoint, Excel, OpenDocument, EPUB, RTF, and any plain text or code file. Screenshots and photos go through the same OCR that powers Capture. The text is extracted on your Mac, so reading costs nothing except on that image path, which is a Vision call like a capture.

It reads the bytes, not the name

A PDF saved as .txt resolves, and so does a file with no extension at all. A scan with no text layer is reported as a scan and pointed at Capture, instead of coming back silently empty.

Ask each run, or pin one file

Two modes per agent. The palette can ask for the file first, or the agent can always read the same document. Picking is free: cancel and nothing is billed.

A workflow can start from a document too

A workflow takes the same trigger, with the same two modes, and then does what a workflow does: up to five steps, ending in Jira, Confluence, Linear, Notion or Slack. The minutes go in as a PDF and the tickets come out filed, with nobody retyping anything in between.

Text back,
or the document itself

Every agent says what it gives you. Most hand back text and paste it where your cursor was. A document agent can hand back the document: the same file, rewritten, with the layout somebody spent an afternoon on still on it.

Your document, edited, design intact

Set an agent to give you the edited file and the same Word document, deck or spreadsheet comes back rewritten: the styles, headings, tables and images are the ones that were already there. Khint writes every byte of structure itself and only ever swaps the text inside a paragraph, so a model can propose words and can never corrupt a file.

PDFs, and the cases it won't fake

A PDF is repainted line by line. Text that won't fit its box, a background that's not one flat color, a character the font can't draw: each is named in the preview rather than quietly approximated.

It hands you a copy,
and shows you every line first

A rewrite you can't inspect is a rewrite you can't ship. Nothing here is written back into the file you gave it.

Security and privacy
Your original is never opened for writing
Every result is a new file, never overwriting an earlier one
A preview names every changed paragraph first
What a PDF can't reproduce is flagged, never faked
The model proposes text, Khint writes the file
A file that changed on disk stops the write
The document's text never reaches your history or the cloud
An edit shows its credit estimate before it runs

Three ways to hand it a file

Point an agent at one, ask for the file back instead of text, or start a whole workflow at a document. Each one is a field in an editor you already use.

Point an agent at a file

Set any agent's input source to Document and it reads the file instead of your selection.

Download for Mac
  • The palette's picker: drop zone, recent files, or browse
  • Ask each run, or pin one file for good
  • Small files are read whole, big ones condensed once
  • The pinned summary refreshes when the file changes

Ask for the file back

Switch what the agent gives you from text to the document itself, rewritten.

See what an agent reads
  • Only on a document agent: the editor refuses the other pairing
  • Word, PowerPoint, Excel and PDF all come back
  • A preview shows the changed paragraphs before you keep it
  • Download it, reveal it in Finder, or find it again in history

Start a workflow from a document

A whole routine can begin at a file and end in the tools your team reads.

See workflows
  • Same two modes as an agent: ask each run, or pinned
  • Up to five steps: agents, writes, a parallel fan-out
  • Each step reads what the step before it produced
  • Canceling the file pick cancels the run, nothing half-done

Common questions about running AI on documents

What file types can a Khint agent read?

PDF, Word, PowerPoint, Excel, OpenDocument, EPUB, RTF, and any plain-text or code file. Pages, Keynote and Numbers documents are read through the PDF preview they carry. Image files (screenshots, photos, scans saved as PNG or JPEG) go through the same OCR that powers Capture. Khint checks the content rather than the extension, so a PDF named .txt or a file with no extension still resolves. Two deliberate gaps, both named rather than silently failed: the 1990s binary .doc and .ppt formats, which you save as .docx or .pptx first, and a scanned PDF with no text layer, which Khint reports as scanned and points at Capture & ask.

Can AI edit a Word document without destroying the formatting?

Yes, and that's the point. Set an agent to give you the edited file and a Word document, slide deck or spreadsheet comes back rewritten with its design intact: same styles, same headings, same tables, same images. The model never sees a file format. It's handed the paragraphs as plain text and returns only the ones it changed; Khint writes every byte of structure itself and swaps the text inside each paragraph, which is why a rewrite can be wrong but can never corrupt the file.

Does it overwrite my original file?

Never. The original isn't even opened for writing. The result is saved as a new file, and a preview shows exactly which paragraphs changed before you download it or reveal it in Finder. Edited documents also stay reachable from your Home history, with Download and Reveal right on the row. If the file changed on disk while the agent was working, Khint refuses to write rather than apply the edits to the wrong paragraphs.

How does an agent get its document?

Two modes, chosen per agent. Ask me each run: the palette opens a picker with a drop zone, your recent files and browse, before anything runs. Or pin one file to the agent so it always reads that document: a living spec, a glossary, a style guide. Picking is free: canceling the picker never costs a run. A workflow takes the same trigger with the same two modes, so a PDF of meeting minutes can go in and filed tickets can come out.

What about PDFs?

A PDF edit is repainted line by line, in place, and Khint is honest about the hard cases: text that won't fit its box, a background that's not one flat color, a character the font can't draw. Each one is named in the preview rather than quietly approximated. Reading a PDF works on anything with a text layer; a scan without one is reported as scanned and pointed at Capture & ask, rather than coming back as an empty document.

What does running an agent on a document cost?

The text is extracted on your Mac, so opening the file costs nothing for every text format: only images go through OCR, which is a Vision call like a capture. A file small enough to send whole is sent whole; a big one is condensed once and that summary is cached against the file's own content, so re-running is instant and a changed file refreshes itself on the next run. Editing is the heavy end, so the palette shows the estimate before it bills: Edit document · ~4 credits.

Try Khint today

Free with 300 credits a month. No credit card. macOS 13+ or Windows 10 22H2+.