It opens on top of the text you selected
Word, Gmail, Notion, your terminal. Highlight a paragraph, press the shortcut, and everything you've saved opens over it, ready to run.
Select text anywhere, press ⌘⇧K, and your own prompt runs on it. The result comes back where you were, or lands in Jira, Notion or Slack.
Hi, our CSV export hangs at 99% since Tuesday. We are blocked on month-end reporting. Can you help?
Client email to documented fixYou're already in the document · nothing has started
Agents
Select text in any app, press ⌘⇧K, and run your own prompts on it. Hand an agent a PDF or a Word file and you get the file back edited, design intact. Bundle your agents into a pack, chain up to five of them, and ship the result to Jira, Notion or Slack.
Word, Gmail, Notion, your terminal. Highlight a paragraph, press the shortcut, and everything you've saved opens over it, ready to run.
Packs swap the whole palette. Your product prompts in the morning, support in the afternoon, nothing else in the way.
One line about what it should do. Khint writes the prompt, names the agent, picks what it reads and finds a shortcut nobody has taken. Nothing starts blank.
A run can file the ticket, post the link and update the page, all on its own. Khint writes to them directly and logs what it did. Nothing to copy over.
An agent can search your Jira, Confluence, Notion and Linear. Indexed on your Mac, never on our servers.
Give an agent its own keyRecord a shortcut and the agent fires without the palette. Khint checks that combo against every binding you already have.
Share a pack with your teamSend one link. They open it, copy the pack, and get the same agents in their own palette.
Local historyEvery run is stored on your device and searchable in full text. Never uploaded.
Idea bar
You won't always know which agent you need, or whether you've ever built one. Say what you want done instead, and let Khint find a route through what you already have.
Hold your dictation key with the palette open and talk. The words land in the bar, where you can still read and fix them.
Every step, the tool it writes to and the text it runs on, before anything fires. If it got you wrong, close the palette and nothing happens.
Khint offers to save it as an agent or a workflow with its own shortcut. A library you described rather than one you built.
Draw a region over a dashboard, a PDF, a screen share. Khint plans from what it read: nothing to select, nothing to retype.
Reading a plan never costs a credit. You only pay for the steps you let run.
It plans from your agentsNot from a generic model menu: your own saved prompts and the tools you connected.
You see it before it runsEvery step, every tool it will write to, and the text it will run on.
Speak it insteadHold your dictation key with the palette open and the idea types itself.
Capture
Draw a region and Khint reads it. Extract text straight to your clipboard, or open a chat and ask questions about what you captured. No app switch, no retyping.
Premium €9.99 → €12.99, live May 15.
Sprint Q2: pricingPremium €9.99 → €12.99, live May 15.- CS owns the FAQ- Comms owner TBD
Capture a region or drop an image in. Both become text you can use.
Feeds your sessionCaptures land in your active Memory session, so later actions see them.
Recent capturesA reviewable feed of everything you've captured, with one-click copy.
Speak it insteadHold a shortcut, talk, and the text lands at your cursor. Agents and workflows can be voice-triggered too.
Memory
Every other AI tool makes you paste the context back in. Khint already has it: what you've been doing since you started, and what your own tools already say about it.
Start a session when the work starts, stop it when it ends. Notes, captures, files and runs land in it as they happen, on your machine and nowhere else.
Khint keeps rewriting the session into the objective, the open threads and the decisions. That's all an agent reads, no matter how long the day was.
Two switches on the agent itself: one for the active session, and one that swaps emails, phone numbers and account numbers for placeholders before the text reaches the model. Your indexed tools have theirs on the Memory page, one per source.
Jira, Confluence, Notion and Linear sync into a local index. An agent that needs the spec finds it in milliseconds, without a model call and without a copy on our servers.
A PDF, a Word file or a deck becomes text in the session, right where you are.
Searching is freeKnowledge runs on your machine, so a lookup never calls a model or spends a credit.
Your other AI sees itClaude Desktop and Codex read the same session over MCP, with no copy to keep in step.
Leave with the recapOne click copies the whole session as a summary you can paste into anything.
KH-184 lands in your session, your agent writes the draft with the full context it needs, and the result goes straight back to Jira.
Agents, Capture, Memory and every tool you connect, in a single window. Click the sidebar to navigate: it's the same UI you get on the desktop.
Khint strips invisible characters out of any text it handles, because they are how a hidden instruction travels inside something that looks ordinary. Two of the characters on that list are not invisible decoration at all: one is what joins a family emoji into a single glyph instead of three separate people, and the other is a letter in Persian and in the Indic scripts. Removing them from text coming in is right: that is where a hidden payload would arrive. Removing them from text going back out was corrupting every emoji sequence and every Persian word an agent produced, on every paste and inside every document Khint edited for you. Those two are left alone on the way out now; everything genuinely invisible still goes, on the paste path, the ask-in-words path and the document path alike. Then the one model answer that reached storage untouched: the running summary of your work session. It is never shown and never pasted, but it is read back into every prompt for the rest of the session, so anything an entry smuggled into it travelled with you all day. It is cleaned now, on both routes that write it. That same summary could also be scheduled for a rebuild and have the rebuild cancelled in the same breath: setting a session objective and the automatic refresh every fifteen entries were both marking it stale in a way a rebuild already under way could wipe, leaving agents reading an older picture than the one you had asked for. Three smaller ones ride along. The little note under a run saying which context was used was worked out by reading your session and your knowledge a second time, so it could describe something other than what was actually sent, and made every run pay for a second search to do it; it now reports exactly what went. Khint's connector for other AI tools now publishes how long an entry may be, instead of letting an assistant discover the ceiling by being refused with the work already done. And the "manage subscription" button in Billing used to print the payment provider's own sentence, customer reference and all, when something was wrong with the account; it says what is actually happening and what to do about it.
AUG 10, 2026Super Khint proposes a list of actions, you run the ones you want, and when you close the card it writes down what happened: a line on your home page and an entry in your work session. There were three ways a run could end and only one of them wrote that record. Close the card while a write was still going out and the write landed with nothing left to attach its receipt to: the ticket existed in Jira, the run left no trace anywhere in Khint. Start a second Super Khint while a card with completed actions was still open and the first run's receipts were overwritten where they stood: closing that first card afterwards then did nothing at all, in silence, for actions that had genuinely written to Jira or Slack. All three endings write the record now, and whoever finishes last owns it, so closing under a slow write waits for the write rather than pulling the run out from under it. The fourth fault is the one that could cost you a duplicate ticket. The panel can rebuild itself from scratch (on Windows the system recycles it), and the rebuilt list came back with no memory of which rows had already run: every line offering a live Run button again, and one press of Run all re-creating tickets that already existed. Rows that succeeded are remembered across a rebuild, and the guard sits at the moment of execution rather than in the display, so a second press hands back the receipt it already has instead of writing a second time. And a question the panel asked you (a Jira key, a page address) that disappeared under one of those rebuilds used to take the keyboard with it, leaving you typing into your own app with the keys going somewhere else, and left Run all stuck for the next plan too. A question that goes away now answers itself with a cancel, which is what it meant.
AUG 10, 2026The palette borrows your clipboard for a moment: it copies whatever you have selected so it can read it, then puts back what was there before. Every way out of the palette owes you that, and four of them had quietly stopped: Capture, the PowerPoint diagram, sending to Claude, and opening another app all left the borrowed copy parked, where a later run with no palette even open could write it over your clipboard. That was one forgotten decision per row, fixed one row at a time while the next one shipped; there is a single rule now, so the row added next is covered by default. PowerPoint is the one deliberate exception, because it reads that copy itself. Three more places had the same shape: a plan started from the ask-in-plain-words bar never restored anything at all, and Super Khint restored only when it had read your selection, never when it read your clipboard or a file, and not at all if you dismissed the file picker. Then the shortcuts. A workflow's own key combination is registered across all your packs on purpose, but pressing it only ever looked inside the pack you had active: if the workflow lived in another one the key did nothing and said nothing, while the very same workflow started by voice worked, because that route looked everywhere. Both look everywhere now, and a workflow that genuinely cannot start says so instead of failing in silence. Two more you would have felt without being able to name. On Windows, any app whose name merely contains "cmd", Total Commander among them, was having text typed into it character by character and losing the first one. And the palette opened on whichever screen it was last on rather than the one your pointer is on, so on a multi-monitor desk it could appear on a display you were not looking at. Finally, the ask-in-words bar and Super Khint were refusing runs that write to Jira or Slack unless a work session was active: a rule that stopped applying weeks ago, that no other route enforced, and that was blocking work which would have gone through perfectly well.
AUG 10, 2026Free with 300 credits a month. No credit card. macOS 13+ or Windows 10 22H2+. Universal binary on Mac.
Yes. The palette opens on an idea bar: type what you want done ("summarize this and file it in Jira") and Khint plans it from your own agents and the tools you connected, shows you the plan and the text it will run on, then runs it when you confirm. Planning is free. You can also speak the idea while the palette is open, or hit "Do this" on a screen capture. If the run was good, Khint offers to save it as a reusable agent or workflow with its own shortcut, so you end up with a library you described rather than one you assembled.
Most AI features only see the text you select right now. That's not enough for real desktop work. You also need what came before. Khint keeps a lightweight session of what you're doing (meeting notes, drafts, code, OCR captures) and auto-attaches a compact summary to every prompt, so your agent sees the full context. The session lives in local SQLite on your device, and is exposed read+write to Claude Desktop over MCP. Your other AI tool sees the same context.
Both live behind a keyboard shortcut, but Khint is AI-first: AI text actions on selected text are the core feature, OCR is bundled, and a Memory session feeds your active work context into every AI prompt. Khint also exposes an MCP server so Claude Desktop can read your active session. Raycast doesn't. Free includes 300 credits a month (about 10 AI actions a day), with all 6 work integrations on every plan. Paid plans start at €7/month billed annually: Pro at €29/month gives 4,000 credits a month, and a one-time €349 Lifetime (bring your own AI key, launch offer) sits above.
Apple Intelligence covers a fixed set of single-shot rewrites in supported apps. Khint runs in any app on Mac or Windows, lets you author your own prompts, bundles them as Packs, and chains up to 5 prompts in series. Plus a screen-to-text OCR shortcut and a Memory session that gives Claude your active work context over MCP.
Only the text you select when you trigger an agent, and only at that moment, is sent to the AI provider that serves your run. Capture reads the screen region on-device, then sends that image for text extraction; nothing is kept on the provider's side beyond the call. Memory sessions live in a local SQLite file. Cloud sync is opt-in via sign-in. See /privacy for the providers we use, and /security for the full data flow.
Yes. Select the text on a slide and run Transform to diagram from the palette. Khint classifies it into one of ten consulting frameworks (2×2 matrix, pyramid, process, waterfall, timeline, comparison and more) and renders a live preview. Switch between 36 layout variants, apply curated palettes, then save a .pptx built from native, editable PowerPoint shapes, or a transparent PNG. macOS only, and it counts as one run.