Khint proposes,you approve every write
A stand-up leaves you with a pile of work that changes shape every day: two tickets to create, one to comment on, a page to append, a message to post. A scripted workflow cannot know that in advance. Select the notes, press ⌘↵ on “Let Khint handle this”, and Khint reads them, proposes each write with its target already filled in, and runs nothing at all until you click Run on that row.
A fixed script cannot
match a variable meeting
A saved workflow is scripted on purpose: when you build it you pick the write and its target, and every run does exactly that. It is the right contract for a job you repeat, like turning notes into one ticket in one project. It is the wrong contract for a meeting, because a meeting produces a batch of actions that varies in number and in kind. Some days it is three tickets. Some days it is two comments on tickets that already exist and a follow-up email. The Khint row in the palette inverts the contract: Khint reads the content and proposes the list, and your control moves from framing the work in advance to approving it before it is written. The writes themselves are the same integration writes every other workflow uses.
The real palette running a saved workflow: the notes leave the document and the writes land in Jira, Confluence and Slack. The Khint row makes the same writes, with an approval row in front of each one.
From raw notes to approved writes
Turning a daily stand-up transcript into the tickets, comments and messages it actually calls for.
Select the notes, open the palette, press ⌘↵
The palette always carries one row at the top: “Let Khint handle this”. Select the notes in your doc, or copy the transcript, open the palette with Cmd+Shift+K and press ⌘↵ on that row, or click it. Its second line already says what it is about to read, the words you selected or what is on your clipboard, and when there is nothing to read it offers you a file instead. Nothing to create first, nothing to set up. It reads the notes in full: a condensed transcript loses the per-ticket detail, and that detail is exactly what has to survive for the proposal to name the right issues.
Aim it once, per job, or not at all
Steering is optional. A few lines in Settings › General, under “How you work”, apply to every run: tickets in French, the default project is KH, never mail the client without asking. For a job you repeat, save a brief: it is an ordinary workflowwith a single “Khint decides” step and its own sentence, something like “project ENG, post the summary in #standup”, and it runs from its own row in the palette. Leave both empty and Khint simply proposes what the notes call for.
Khint proposes a list, not a paragraph
One model call reads the notes and returns a list of actions across the integrations you have connected, and only those: create a Jira issue, comment on one, append to a Confluence page, add a Notion page, open a Linear issue, post to Slack, draft a Gmail message, or simply put text on your clipboard. A paused integration is invisible to it. If you have turned on Khint's local Knowledge index, Khint also sees excerpts from your own Jira and Confluence content, which is what lets it comment on the ticket you already have instead of creating a second one.
Approve each action in the panel
When the plan is ready, a panel titled Khint opens with one row per proposed action: which tool, which target, and a preview of what would be written. Each row has its own Run button, with a quieter Run all beside them, and nothing happens until you press one. Run the first, read the receipt with the key or the URL that came back, then decide about the second. Where Khint cannot know a value, a Jira issue type, a Confluence space, a Linear team, the row asks you with the same picker the rest of the palette uses. A failed row offers Retry and leaves the rows around it alone.
Close the run and keep the trail
Skip the rows you disagree with, they simply never run. When you close the panel, Khint records the run once in your history, as “Khint · 3 actions done”, and each write that landed has already logged an entry against your active Memory session with the link back to the ticket or page. The next morning you press ⌘↵ on a new transcript and get a different list, because the list comes from the content and not from a script.
Two honest ways to automate a meeting
Both ship in Khint. Pick by whether the work repeats identically or changes with the content.
A scripted workflow
You fix the steps and the targets when you save it, and every run does exactly that.
How workflows work- Best when the job repeats: the same ticket in the same project, every time
- Runs on your own keyboard shortcut, with no window to look at
- Can fan out to several tools at once, all receiving the same text
- Cannot add a write the content asks for, because it was never in the script
Let Khint handle this
Khint reads the content and proposes the list. You approve it row by row.
Download Khint Free- Best when the batch changes every time: a stand-up, a discovery call, a review
- One permanent row in the palette, on ⌘↵, for whatever you are looking at
- Can comment on tickets you already have, not only create new ones
- Nothing is written until you click Run on that specific row
What it refuses to do
The boundaries are the product. An agent that quietly widened them would not be worth approving.
No unattended mode
There is no setting that runs the proposed list for you. Nothing runs without your Run, one row at a time or Run all once you have read the list, and a plan you close early simply stops there.
No global shortcut
The row answers to ⌘↵ once the palette is open, not to a shortcut of its own, because the approval panel has to open somewhere you can see it. A saved brief has no keyboard shortcut either: it runs from its row in the palette.
It is the whole workflow
A saved brief holds exactly one step, Khint decides, so it does not sit in the middle of a longer chain. Scripted steps and that step stay separate workflows.
Nothing is dropped in silence
An action naming a tool you have not connected, or one Khint got wrong, appears in the panel as a line you can read. It is never quietly removed from the list.
What landed stays landed
If one approved write fails, the ones that already succeeded are not rolled back. A message posted to Slack cannot be un-posted, so Khint reports the failure instead of pretending it can undo it.
One billed action per plan
The proposal counts as one AI action. The writes you approve call no model at all, so approving five of them costs nothing beyond that first call.
The approval panel is also the security boundary
Notes you paste are untrusted text, and a transcript can contain a sentence written to steer a model. Khint wraps that content as data before the model sees it, but the stronger defence is the one you can see: an action nobody in the meeting asked for shows up as a row in the panel, and you do not press its Run button. That is the practical difference between an agent with tool access and an agent with tool access plus a person. If the notes themselves are sensitive, you can also redact PII before the call, or read how integrations keep their credentials in your system keychain.
Common questions
What does human-in-the-loop actually mean here?
It means no write reaches one of your tools without a person pressing a button for that specific write. Khint splits the job in two: it reads your notes and proposes a list of actions with their targets filled in, then a panel opens with that list and one Run button per row. Reading and proposing are automatic. Writing is not. You can run three of the six proposed actions, change a target, retry a failure, and close the panel on the rest.
How is this different from a normal Khint workflow?
A normal workflow is scripted. When you build it you choose each step, each integration write and its target, and every run repeats that exactly. It is the better tool when the job is identical every time. The Khint row is for the opposite case: Khint decides how many actions there are and what they point at, based on the content you gave it. One stand-up might produce two tickets, another might produce a comment on an existing ticket and a Slack post. A scripted chain cannot represent that, and a script that guessed would be worse than one that asked.
Can I steer it without writing a brief every time?
Yes, in two layers. Settings › General has a field named How you work: a few standing lines, tickets in French, default project KH, never mail the client without asking, applied to every run ahead of anything else. For a job you repeat, save a brief: an ordinary workflow with one Khint decides step and its own sentence, which runs from its own row in the palette and gets the last word on that job. Both are optional, and with both empty Khint still proposes what the notes call for.
Can it update tickets that already exist, or only create new ones?
It can do both, as long as it can identify the existing item. That is what Khint's local Knowledge index is for: it syncs your Jira, Confluence, Notion and Linear content into a search index on your own machine, and Khint gets the most relevant excerpts for your notes. With it turned on, a mention of the work you discussed can resolve to a real issue key and become a comment on that issue. Without it, Khint has no way to know the item exists and will lean towards creating.
What happens if Khint proposes something wrong?
You do not run that row, and nothing happens. If the target is wrong rather than the action, you can change it in the panel with the same picker the rest of the palette uses, which is also how you fill in values the model cannot know, like a Jira issue type or a Confluence space. If a row names an integration you have not connected, it appears as a visible unmet line rather than disappearing, so you always see the full proposal and not a filtered version of it.
Does my meeting transcript leave my machine?
The transcript is sent to the model for the single call that produces the proposal, the same way any AI action sends the text you selected. It is not stored on Khint's servers as a document, your Memory sessions stay in a local database on your machine and are never synced, and the Knowledge index is built and searched entirely on your Mac or PC.
Let Khint propose, keep the last word
Free with 300 credits a month, about 10 AI actions a day. No credit card. Connect a tool, select your notes, press ⌘↵, and approve the writes one at a time.