One AI toolkit fordata analysts
Khint does not touch your warehouse. It takes the other half of the job: questioning the chart on your screen, turning findings into something a stakeholder reads, and answering the ticket that asked for the number.
The query took twenty minutes,
the write-up takes the afternoon
A chat tab cannot see the dashboard you are staring at, the ticket that asked for the number, or the four things you already ruled out this morning. So every answer starts with you retyping the context. Khint works on whatever is on screen and keeps the rest of the thread in a work session that stays on your machine.
Confluence · Team Velocity Gadget
The real Capture and ask window, questioning a chart that sits in a tool with no export button.
One ad-hoc request,
question to write-up
Every step below is a shipped feature, not a roadmap slide.
- 09:15Open a session for the question you were askedPress Start on a Memory session named after the request: 'Q3 renewals: why the dip'. From here every note, capture and action is folded into a short working state that later actions read. It lives in a local SQLite file and is never synced to a server.
- 09:40Find how the metric was defined last timeKhint keeps an index of your Jira, Confluence, Notion and Linear content on your own machine. Tick Search Knowledge on an agent and the run quietly pulls the matching excerpts, so the definition your team agreed on in March comes back without a tab hunt.
- 11:00Question the chart instead of retyping itThe number lives in a BI tool with no export. Open the palette, pick Capture and ask, drag a region over the chart, then ask about it across several turns on the same screenshot. If you only want the table underneath, Extract text puts it straight on your clipboard.
- 14:20Findings to something a stakeholder readsSelect your rough notes in any app, hit Cmd+Shift+K and run a Write the summary agent. Because the session is active, the draft already knows what was asked, what you captured and which caveats you flagged, so you are editing a draft rather than starting one.
- 15:30File it in both places people lookOne workflow: an agent formats the write-up, then a parallel step posts it to a Slack channel and appends it to the Confluence page in the same run. A separate step can drop the answer as a comment on the Jira issue that requested it.
- 17:00Hand the whole thread to your AI clientClaude Desktop and Codex read the live session over the MCP server that ships with Khint. Ask what you already ruled out today and get an answer built on your own notes, with nothing pasted back in.
Start from a pack,
not a blank prompt
No pack ships pre-tuned for analytics, and pretending otherwise would waste your first evening. What ships is a starter gallery you can strip for parts: Doc Polish for the write-up, Meeting to Ticket for the requests that arrive verbally. Add one, then rewrite the prompts in your own words on the Agents page. If you would rather not write a prompt at all, describe the agent in a sentence and press Generate. Describe a whole sequence instead and the workflow generator assembles the steps. When one earns its place, share the pack and the analysts around you clone it.
The four pieces of the toolkit
Product overviewCommon questions from data and BI analysts
Does Khint connect to my database or write SQL?
No, and that is the honest boundary. Khint has no connection to a warehouse, a BI tool or a query engine, and it never runs a query. It works on what you hand it: the text you selected, a region of your screen, a file you point an agent at, or your voice. The querying stays in your usual tools; Khint takes the reading, questioning and writing that happens after the numbers land.
How does it read a dashboard I cannot export?
Drag a box over it. Extract text runs OCR on that region and puts the result on your clipboard, which covers the table a BI tool refuses to export. Capture and ask instead opens a chat on the screenshot, so you can ask what the trend does after the third sprint and keep asking on the same image across several turns. Nothing is sent anywhere until you trigger the run.
Can it read an exported spreadsheet?
It can read one as text. An agent whose input is set to Document extracts the text of a .xlsx, .docx, .pptx, PDF or plain export locally, and only the extracted text is used. It is not computing over your cells: treat it as a way to summarise, restructure or write up a sheet, not as a calculation engine. Documents with an edit-in-place agent come back as the same file with its formatting intact.
Do the numbers I work on leave my machine?
Only what you explicitly act on. An action sees the text you selected plus, if you switched it on, the active session's compact summary. Memory sessions and the local Knowledge index live in a SQLite file on your machine and are never synced. Any agent can have Redact PII switched on, which replaces emails, phone numbers and account identifiers with placeholders before the text is sent. Integration credentials stay in your OS keychain.
Which tools can it write into, and is there a free plan?
Connect Jira and Confluence with an email plus API token, Linear or Notion with a key, Slack with a token, Gmail through sign-in. The palette then gains rows like Comment on Jira issue, Append to Confluence page and Post to Slack, and a workflow can end on any of them. The free tier is 300 credits a month, roughly 10 actions a day, no credit card. Paid plans start at €7 a month, and every plan includes all the work integrations.
Write up tomorrow's analysis with it
Free with 300 credits a month, about 10 AI actions a day. No credit card, no warehouse connection to approve: it reads your screen and your selection, nothing else.