Your AI already readyour Jira and Confluence

Khint keeps a full-text index of the work tools you connect on your own machine. Agents search it before they draft, so the ticket you get back is written in your product's vocabulary, against your real backlog, with nothing uploaded anywhere.

Indexed on your machine, never on our servers · Zero credits to index and to search · Jira, Confluence, Notion, Linear

A general AI writes confidently
about a product it has never seen

It does not know that the thing you call an export is called a data extract in your backlog. It does not know a ticket for this already exists. So you spend the first five minutes of every draft pasting the context back in, and the next twenty hunting for the page that would have made the draft right.

The usual fix is to hand your whole workspace to a cloud platform and let it keep a synced copy of your Confluence forever. Khint takes the other route: Knowledge builds the index on your own machine and leaves it there. Retrieval is a local search, so it costs no credits and answers in milliseconds, and the only thing that ever reaches the model is the handful of passages that actually match the run you just launched.

Three toggles, then it is just on

Setup is a one-time thing, and every step is a screen that ships today.

1Connect the tool onceOn the Integrations page, connect Jira, Confluence, Notion or Linear with an API token. Credentials go into your OS keychain, not to a Khint server, and you connect with your own account: what your Jira cannot show you, your index will never hold.
2Switch the source on under MemoryOpen Memory and find the Knowledge section under your session list. Each connected source has its own toggle. Flip one on and Khint starts indexing it on your machine in the background, then keeps it fresh incrementally every 30 minutes or on demand with Sync now.
3Let the agents that need it searchOpen any agent in the editor and switch on Search Knowledge. From then on, every run of that agent first searches your local index for the passages that match the text you are acting on, and hands them to the model along with your selection.

One refinement week,
with the backlog in the room

A product owner, a request that sounds familiar, and an index that already holds the answer.

  1. Tue 09:40A request lands, and it sounds familiarA stakeholder asks for a CSV export on the reporting screen. You are fairly sure this was discussed two quarters ago, in a ticket nobody can name and a Confluence page nobody can find. The usual move is to ask in the team channel and wait.
  2. 09:41Ask the index instead of asking the teamMemory has a Test the index bar under Knowledge, labelled see exactly what your agents will read. Type the request into it and the matching passages come back from your indexed Jira and Confluence with the matched words highlighted in place. You get the prior art yourself, in seconds, without interrupting anyone.
  3. 09:45Draft the story with your backlog in the roomSelect the request and run your user-story agent, the one with Search Knowledge switched on. The closest excerpts from your real tickets and spec pages travel with the selection, so the draft lands in your product's own vocabulary instead of generic SaaS wording. The confirmation reads knowledge context used, so you always know when a run was grounded.
  4. 09:47File it without leaving the keyboardCreate the Jira issue straight from the palette. The picker opens before the model call, so you choose the project first and nothing is spent if you change your mind. The write lands as an entry in your active Memory session, which is what makes the next draft factual rather than remembered.
  5. Wed 11:00Refinement, without the tab huntGrooming the same story, run your acceptance-criteria agent on it. The spec pages that describe how exports already behave elsewhere in the product come along automatically, so the criteria you write are consistent with what is built rather than with what you assumed at 9am.
  6. Thu 14:20A new joiner asks why it works that wayThe decision pages are in the index too. Run your explain agent on their question and the answer is assembled from the passages your own workspace holds, so you can point them at the real page rather than reconstructing the reasoning from memory.

What you did today,
and what your company already knows

Khint carries two different kinds of context, and it is worth keeping them apart. A Memory session is the short-term one: what you pulled, drafted and filed since you pressed Start this morning, compacted into a small working state that agents can read. Knowledge is the long-term one: the tickets and pages your organisation wrote long before today. They are independent opt-ins on an agent, and the run confirmation names whichever ones were used, so you are never guessing what the model saw. An agent that turns a stand-up dump into tickets wants the session. An agent that checks whether a request is already specified wants Knowledge. Plenty of agents want both.

What it is, and what it is not

Retrieval you cannot predict is retrieval you stop trusting. The boundaries, stated plainly.

Four sources, not everythingJira, Confluence, Notion and Linear. Your email, your drive and your chat history are not indexed.
Keyword ranking, not meaningBM25 full-text search over the terms in your text. Fast, local and free, but a pure synonym can miss.
A working set, not an archiveUp to 2 000 of the most recently updated documents per source, so the index stays about live work.
A few excerpts, not the corpusOnly the top matches for that one run reach the model, capped and sanitized before they are sent.
Scheduled pull, not a live mirrorA background pass every 30 minutes, plus Sync now. Nothing is watched and no webhook is involved.
Opt-in per agent, not always onAn agent searches Knowledge only if you switched it on, so your simple text agents stay untouched.

The index stays put,
the excerpts travel

Everything Khint indexes lives in a database on your own machine. There is no copy of your corpus on a Khint server, no embedding service in the middle, and nothing to switch off at the vendor if you change your mind: turning a source off, or using Forget credentials on its integration card, deletes that source's local index immediately. Because you connect with your own account, your index holds exactly what your account can already see, so there is no second permission model to get wrong. When a run does opt in, only the matching excerpts are sent, capped and sanitized first, and wrapped in a tagged data block that the model is instructed to treat as untrusted reference material rather than as instructions. If the text itself is sensitive, the same agent can also switch on PII redaction, or you can run your own API key.

Common questions about the local index

Does Khint upload my Jira and Confluence anywhere?

No. The index is a full-text database that lives on your own machine, and Khint keeps no copy of your corpus on its servers. Your Jira, Confluence, Notion and Linear credentials stay in the OS keychain. The only thing that ever travels is the handful of matched excerpts that ride along with an agent run you launched yourself, sent inside a tagged data block the same way your active session summary travels today. Switch a source off, or use Forget credentials on its integration card, and its local index is deleted straight away.

Is this semantic search, or does it use embeddings?

Neither. Retrieval is local BM25 full-text search: Khint pulls the meaningful terms out of the text you are acting on, drops the French and English stopwords, and ranks your indexed documents against them. There is no embedding model and no vector service, which is exactly why it can run on your machine, answer in milliseconds and cost nothing. The practical consequence is worth knowing: matching follows the words in your text, so a request that shares vocabulary with the ticket you are hunting will find it, while a pure synonym may not.

Does Knowledge cost credits or slow my actions down?

No on both counts. Indexing makes no AI calls at all, and the search that runs before an agent is local, so it adds milliseconds rather than a round trip. The only cost is the few hundred extra words of excerpt added to the runs that opted in, which the normal per-run rounding usually absorbs. Retrieval is also per-agent opt-in: an agent only searches the index if you switched Search Knowledge on in its editor, so your purely textual agents stay exactly as cheap and as fast as before.

How fresh is the index, and how much does it hold?

Khint runs a first pass shortly after launch and then syncs every 30 minutes in the background, and there is a Sync now button when you do not want to wait. Sync is incremental: after the first pass only the documents that changed since the last run are fetched. Each source keeps up to 2 000 of its most recently updated documents, so the index stays a working set of live work rather than a full archive. It is a pull on a schedule, not a live mirror: nothing is watched and no webhook is involved.

Index your workspace, keep it yours

Free with 300 credits a month, about 10 AI actions a day. No credit card. Indexing and searching cost nothing, and connecting a tool takes an API token, not an admin.