killBottleneck is in public beta — cloud and self-host.killBottleneck is in beta.Beta on GitHub →
killBottleneck

English

Čeština

English

Čeština

Appearance

Try it freeTry free

The AI assistant on the side ​

Since v0.61-beta killBottleneck has an AI assistant in a panel on the right. It is with you on every page of the full app — Tasks, My day and the map editor — and the conversation is not lost when you move between pages. The same rule applies as for the rest of the AI: it may read right away, but changes things only after you confirm.

The assistant lives in the panel on the right; in the map editor it also offers the Break down this project chip.

Where to find it ​

  • The top bar has a robot button AI assistant; the map editor has an Assistant button. Both open and close the panel.

  • A closed panel stays as an Assistant tab on the right edge of the screen.

  • Drag the panel's left edge to change its width (320–640 px). Minimise (the conversation stays) hides the panel while the conversation carries on.

  • On a narrow display in the full app, the minimised assistant turns into a bar at the bottom: you see the last sentence of the reply and can keep typing without opening the panel. A dot in the bar means a question or a confirmation card is waiting for you.

  • In the simplified view on a phone, the header has an Assistant button: it opens the panel full-screen, the down arrow closes it. It is the same conversation as on the computer. The Show in map link on a card opens the map in the full app.

The Assistant button is always in the header — even on an instance with no AI configured, where the assistant answers from the help, offers Documents and Relax, but does nothing in the maps (see When the AI is down).

What the assistant can do ​

An empty conversation starts with "I can see your projects, idea buffer and rules. Changes happen only after you confirm them." and a row of chips: Morning briefing, Break down a project (in an open map Break down this project), Go through my idea buffer, What is on my plate today?, Where is my bottleneck? and Switch my look. Click one or type your own sentence.

Reads right away ​

The assistant opens whatever it needs for the answer: the project list and maps, My day, the idea buffer, rules and rule templates, the Organization overview, the list of people and your calendar events ("what meetings do I have this week?"). A small Looked at: … line above the reply tells you what it based the answer on.

While you wait you do not stare at a spinner either: the Thinking… bubble keeps saying what the assistant has just done and is doing — e.g. Read: project map "Workshop" · Thinking…, Reading: My day… or Preparing a proposal to confirm…. Most useful with a slower model on your own server, where the first reply after the model loads can take a minute.

It searches the archive too. Every project has its sequential number (#12 — shown in the project dashboard, in the tooltip of the map title and in the Archive; the assistant tells you too) and the assistant searches all your projects, archived ones included: "Where did we deal with the fencing invoice?", "What was in last year's kitchen project?" or simply "Open #12". It finds a project by number, title and by the titles, descriptions and owners of its steps and then reads an archived project just like an active one. A few key words in any form are enough ("invoices" finds "Invoice"), not all of them need to match — results are ranked by how well they fit and rare words weigh more. It also searches your Documents (drafts, meeting notes, prep notes). It only sees the projects you can see, and only your own documents.

Where it is stuck — risks right in the map ​

Ask in an open project "Where is the bottleneck?", "What is overdue?" or "What is stuck here?" and the assistant lists the steps that are overdue, the ones blocking open sub-steps, the ones waiting for their sub-steps and the ones idle for 14+ days (the same definition as My day and Organization — no scoring, just facts with day counts). Under the reply there is a Highlight in the map card: the steps get a temporary colour in the map (red overdue, orange blocking, purple idle, amber waiting) and a bar at the top of the map shows the counts with a Clear highlight button. It is only a view in your browser — nothing is written to the map and the highlight goes away by itself.

Changes only through a card ​

Anything that changes your projects arrives as a card describing the change, with Yes, do it and No buttons. Nothing happens until you confirm. This covers:

  • creating a new project (also from ideas already in the idea buffer),
  • adding nodes to a map and moving an idea from the buffer into a project,
  • changing a node — title, status, description, owner, plan ("when I will work on it") and deadline,
  • putting several items into the idea buffer at once (all of them on one card),
  • creating an automation rule and enabling or disabling it,
  • creating an event in the calendar and setting a deadline reminder on a task.

When something goes somewhere, the card draws it: for an idea from the buffer, a moved step or steps added under a step you see where it comes from on top, then the project and the parent steps, and the new or moved row highlighted. Steps and projects get icons from the assistant (an emoji that fits the name) right when they are created; say "add icons everywhere" and it proposes them for the whole project in one card. A longer answer is structured into blocks — a part starting with a label (Done, On fire, Recommendation…) has an icon and a coloured edge.

After confirmation the card switches to Done and offers Open project or Show in map (the latter also at the very bottom of the answer when more text follows the card). If you decline, the card shows Declined and the assistant looks for another way. Changes go through the same interface as the rest of the app, so your permissions apply: the assistant cannot do anything you could not do yourself, and notifications go out as usual.

Without a card the assistant only does small things that do not change your projects: note a single idea into the buffer ("Into the idea buffer: …"), save a note to its memory, switch the look (with a Revert button) or light/dark mode, change your personal settings (language, map readability, notifications — always with Revert, see Settings) and prepare a text for you to copy.

Nothing happens until you click Yes, do it. The card shows the steps, their deadlines and who they will be assigned to.

Short questions with prepared answers ​

Below every reply there is always something to click: a question with options, a card to confirm or "what next" chips. When the assistant is only waiting (for a photo of your notes, say), a chip offers the other choice — Nothing to add, go on; elsewhere at least What next?

The opening step of the guided flows (the invitation to send a photo, a voice note or ideas, the new project form) comes from the app itself, so it is there at once, without waiting for the AI. Tap I will write my ideas (or I will send a photo…) and the app just replies Go ahead and waits for your material — the AI starts working once you send it or answer.

When a request is ambiguous, the assistant asks — at most three questions, each with two to four clickable answers. The block starts with "Questions to clarify (N) – pick an answer or write your own:"; instead of a chip you can type an Own answer. Send with Send answers or with Enter in the own-answer field (with several questions Enter first jumps to the next unanswered one).

At the end of a reply it usually offers "what next" chips too — a click sends them as your message. When you describe a goal or a problem ("I'd like to work with fewer interruptions"), the options always include something like "Advise me how to do it": you get 3–5 concrete steps tied to your map and an offer to write them in as sub-steps.

For steps with a deadline the assistant always asks whether you handle them — otherwise they would be missing from My day.

A new project from scratch ​

Type, for example, "I'd like a new map for a hot-dog stand". The assistant proposes the title, the goal and 5–8 first steps itself and offers them on one card. You are always the owner — it does not ask for an owner e-mail. The card shows the whole tree that will be created — indented steps with owner, deadline and plan (for a longer proposal the first 12 and a Show the whole tree button), so you know exactly what you are approving. After creation it offers fitting preparations (a financial overview, suppliers…).

When you want to think the plan through more carefully, use New project with AI — it asks first and then proposes the whole tree at the level of detail you choose.

After confirming: Done, a way into the new project and suggestions for what to do next.

Breaking down the selected node ​

In the map editor the assistant knows which node you have selected. Click a step in the map and type "break this step down" — "this" is the selected node. The proposed sub-steps arrive as a card.

The Break down this project chip goes through the whole open map and recommends what to do with it next.

Plans, deadlines and reminders ​

  • Plan — when you say "I'll do it tomorrow" or "on Monday" about a task, the assistant proposes planning it for that day. A plan never moves the deadline; the map shows a small Plan badge next to the deadline badge.
  • A deadline is a date agreed with someone else (a meeting, a delivery, a hand-over). When one follows from the conversation or a pasted e-mail, the assistant proposes to set, change or remove it — always on a card showing the old and the new date.
  • A timed deadline reminder — say "remind me about the Novak quote the day before the deadline at 9" and the assistant offers a card Remind about "Novak quote" … the day before the deadline at 09:00 — deadline unchanged. This creates a deadline reminder for you only, the same one you would set in the task detail. The task must have a deadline — if it has none, the assistant first offers a card to set one and only then the reminder. A reminder that could no longer arrive (a time in the past) is refused with the reason. When exactly it arrives is taken from the server's answer, not guessed.
  • Untimed alerts ("let me know three days before the deadline", or for a whole project) are still created as an automation rule, again through a card. The assistant rejects a rule that would never fire — on a node without a deadline, with a reminder day already in the past, or notifying the owner of a node that has none — and explains why.

Calendar events ​

A meeting, the dentist or a video call is not a task and belongs to no project — it is an event in the calendar. The assistant creates one from a single sentence:

"Dentist tomorrow at 2 pm, remind me half an hour before.""Video call with the supplier on Friday at 10, invite Jane."

You get a card Create the event "Dentist" 2026-09-20 14:00 · remind 30 min before; with invitees the card lists their e-mails. Only members of your instance can be invited — the assistant looks the name up in the list of people and, if it cannot find it, says so before the card, not after. When the day or the time is missing, it asks. After confirmation the event is in your calendar and in the invitees', and the reminder arrives in the bell and by e-mail in the instance's time zone.

The assistant also changes or cancels an existing event: "Remind me an hour before the dentist", "Move the dentist to 3:30", "Invite Petr to the meeting too", "Cancel the dentist". Always through a card; an invitee cannot change someone else's event, only leave it.

An existing event can be changed — for example a reminder added later.
One sentence → an event card with the day, time, reminder and invitees. Nothing is created until you confirm.

Asked "what do I have tomorrow?" or "what meetings do I have this week?", the assistant lists the calendar events — your own and those you were invited to.

Team calendar ​

The team's shared calendars — company events, meeting room, time off — are handled by the assistant just like personal events; just say where it belongs:

"Book the meeting room tomorrow at 10 for a client meeting." — when it is taken, the assistant says who has it and offers the next free slot. "The Christmas party is on Friday at 6 pm, put it in the company calendar.""I'm on vacation for a week from Monday." · "Jane is off next week." (for a colleague: admin or manager only) "Who is away next week?" · "What's in the company calendar this week?"

The assistant takes calendar names from what you have created ("which shared calendars do we have?"); a new shared calendar is created only by an admin or manager ("create a calendar Car"), a member is pointed to an admin.

When the time-off calendar has approval switched on, an admin or manager clicks Review with the assistant in the request's detail: the assistant finds out who else is away then, what collides in the company calendar and which of the requester's work is due in the period, writes a summary with a recommendation and offers Approve / Decline / Discuss. The decision is written only after the card is confirmed, and the requester is notified — with the reason, if you give one.

From an invitation to the calendar ​

A lecture schedule on a door, an invitation to a series of seminars, kids' training for the whole term — take a photo of the paper, attach the PDF or paste the e-mail and the assistant lists every date:

(photo of the schedule) · "Here's the seminar invitation, put it in my calendar.""Football for Tom: every Tuesday and Thursday 16:00–17:30 at the stadium until the end of January, except the Christmas break."

  • All dates come on one card — each line with the weekday, time and place. That is where you spot a misread date before anything is saved.
  • Repetition (every Tuesday) is expanded into single days (at most 30); days the invitation excludes (holidays) are left out, and so are dates that have already passed; should one stay on the card, it is marked "already passed" — confirming adds it too.
  • A date that is already in the calendar (same title, day and time — e.g. when you send the invitation twice) is struck through on the card as "already in the calendar" and is not added again. When you have something else at that time, the card says "at that time you have: …" — you can still add it.
  • When the year is missing or a date is illegible, the assistant asks first — it never invents dates. If you have shared calendars, it asks where the events belong (personal, company…).
  • Everything or nothing is written — after you confirm the card. Right after that the app offers Also create a project for the preparation: a project with what needs arranging (sign-up, payment, equipment, preparing for the first session) and the deadlines the invitation states. The event dates themselves are not copied into the project — they stay in the calendar.

Events from one invitation form a series: the detail of any of them has Delete the whole series.

Images ​

The assistant can read an image too: a screenshot, a photo of handwritten notes, a task list from your phone or an e-mail. Add the image in one of three ways:

  • Ctrl+V into the message field,
  • drop it on the field (you will see "Drop the image or PDF here"),
  • the image button Add an image or a PDF next to the field.

PNG, JPEG and WebP are accepted; the browser shrinks the image before sending. While it reads you see "Reading the image…". The assistant then offers what to do with the items: put them into the idea buffer, into an existing project, or create a new project from them (that option is always offered). Several items for the buffer go on one card listing all of them.

The original image is not stored. The conversation keeps the transcript (the Read from the image block) and a small thumbnail; thumbnails are kept only for the last three images in a conversation, older ones say "The image is no longer stored, only the transcript remains."

When the transcript is wrong, fix it with the pencil in the transcript block: edit the text, Send the correction, and the assistant answers again from the corrected text (proposals from the wrong version disappear; it counts as another turn). The pencil is only on the last photo or voice note and only until you have confirmed something from its answer or something was saved from it (a document, a note) — after that, send the correction as a normal message.

The assistant treats text read from an image as data, not as instructions: if it wanted to save something to its memory or note an idea in such a turn, you confirm that on a card as well.

When images do not work

"Reading images is not enabled on this instance — please tell your administrator." means the instance has no model that can read images. The text chat keeps working.

A photo of meeting notes: the conversation keeps a thumbnail and the transcript, the original is not stored.

Voice notes ​

Instead of typing you can record a voice note for the assistant — on the way back from a meeting, or in the evening during evening planning. The microphone button sits next to the paperclip (on a phone, in the bottom bar in place of Send while the field is empty).

  1. Tap the microphone — the first time, the browser asks for microphone permission.
  2. The field turns into a bar with the time (0:42 / 5:00), Cancel and Send.
  3. After Send the voice note goes out right away (like on WhatsApp): the app transcribes it and the assistant treats it like a photo of notes — it sorts the ideas and tasks and offers changes on cards to confirm.

The conversation shows 🎤 Voice note 0:42 with the transcript below (a long one is collapsed). When the transcript gets something wrong (a name, a number, a date), fix it with the pencil in the bubble — the same as a photo transcript above. The recording is not stored — it goes to the transcriber once and the text stays in the conversation. A voice note is at most 5 minutes (then recording stops and is sent by itself); anything under 1 second is not sent.

A finished recording (e.g. a WhatsApp voice note, .opus, .m4a, .mp3, up to 3 MB) can be sent with the paperclip or dropped on the field — it goes out the same way.

When a voice note does not go out (transcription is down, the connection drops), the recording is not lost: the bar The voice note was not sent offers Try again, Download and Discard. The same happens when you close the panel or switch to PDF while recording — recording stops and the note waits in the bar; it is not sent by itself.

When the microphone is missing or does not work

  • The microphone button only appears where the instance can transcribe speech (always in the cloud; self-hosted see below).
  • The microphone works only over a secure connection (https) or on localhost — on an instance opened via http:// on the local network the browser does not allow it.
  • "The browser blocked the microphone" = you declined the permission; allow it again in the site settings (the lock icon next to the address).

Transcription is automatic: names, numbers and dates are sometimes garbled — the assistant asks about unclear ones, and nothing is written without your confirmation.

PDF: merge, split, correct ​

The assistant panel has a PDF tab (document icon in its header). Pick or drop files and merge them in the given order, split by pages (1-3, 5, 8-; with nothing entered every page separately), extract chosen pages into a new file or remove them. All of this happens only in your browser — the file never leaves your computer, it costs nothing and does not count towards AI limits.

The PDF tab: files in the order they will be merged; a page range for splitting, extracting or removing.

Correct text with the assistant reads the page text and hands it to the assistant (the file stays with you). Say what to change — "change the price to 13 200 Kč", "Novák to Nováková", "replace this sentence with that one" — and the assistant proposes a list of replacements on a card (page, text to find, new text). After your confirmation the browser edits the PDF, shows a preview of the page and offers the corrected file for download.

Correcting a quote: counter length and validity — after confirmation a preview of the corrected page and the download.

Honestly, what it can and cannot do:

  • The correction is an overlay: the original spot is covered with the background colour and the new text is written in a built-in font (Liberation, with diacritics). The font may differ from the original; longer text is shrunk (by 15 % at most) or the rest of the line is redrawn, otherwise the assistant tells you it does not fit.
  • The original text stays in the file under the overlay — search still finds it. This is not anonymisation.
  • Scans without a text layer and password-protected PDFs cannot be corrected; at most 60 pages / 40 000 characters of text per attachment (split a longer document first or extract the pages you need).
  • The assistant only receives the page text, never the file; a turn with a PDF counts as two towards the hourly limit.

Ready-to-use drafts ​

Ask for an e-mail, meeting points or call points, a note or a summary and the assistant prepares them in a box with a Copy button. For an e-mail it shows the subject separately. The draft is also saved to Documents (link Open beside); when it belongs to a project, the document links to its map.

Invoices for clients ​

The assistant can issue an invoice to your client — a sentence like “invoice Novák for 12 hours of consulting” produces a card with items and totals, and after confirmation a PDF with a QR payment. Invoices live in Documents under the Invoices filter, the supplier details are set by an administrator in Organization settings → Issuing invoices, and clients are kept in Worked time → Clients. The Invoices page describes it all.

Documents ​

Everything the assistant writes (e-mails, notes, meeting points and summaries) is saved to Documents automatically — like the plan or text other AI tools show next to the chat. Open them with the folder button in the assistant header or with Open beside under a draft. On a computer the panel opens to the left of the chat, over the map (the map does not shrink); when the assistant writes or updates a document in the conversation, the panel opens on it by itself. On a phone the documents cover the whole screen and Back to chat returns to the conversation.

  • The list has a filter All · Notes · E-mails · Summaries · Invoices and search in title and text.
  • Edit (or a double click into the text) opens the text for editing; for an e-mail also To and Subject. Unsaved changes are never lost silently — the panel asks before you leave.
  • Copy, Open in e-mail (opens your mail program with recipient, subject and text filled in — the app sends nothing) and Download as text.
  • Restore previous version — every change (yours or the assistant's) keeps the version before it.
  • You can also ask the assistant to change a document: "make that e-mail more formal" — it finds it and rewrites it (the previous version stays restorable).

The list starts with the pinned assistant memory (see Memory). Only the assistant creates documents; you edit them. Quick ideas and tasks belong in the idea buffer — from there they reach the plan, and the morning briefing and evening planning sort them. When the assistant finds tasks in a document, it offers to put them into the buffer or a project.

Documents are private — of all users only you see them, even in a team instance (technically the server administrator can read them, just like assistant conversations). Up to 200 per user; one holds at most 20,000 characters. They are included in Download all my data.

Pages ​

Need a presentation for management, an overview for a customer or a small interactive page (a calculator, a timeline, a filterable table)? Tell the assistant "make me a presentation of the Q4 results for management" or "build a customer overview from this map". Or skip the wording: in the Pages tab click New page (the same choice sits among the AI helpers and in the chips under the chat) and pick a template — presentation for management, team overview, customer summary, timeline, status report with a traffic light, checklist, or a custom brief — plus the project and the look; the app composes the brief for the assistant. The assistant reads the project, hands over the brief and the app writes the page after its reply. It writes section by section (outline → look and skeleton → each section): the panel lists the steps with the running one, the preview grows and a Code toggle shows the HTML as it comes. It takes a minute or more; keep chatting meanwhile — the finished page shows up in the Pages tab (next to Documents) by itself.

  • Preview right in the panel; Open shows it full-window in a new tab.
  • Share by link creates a secret link (…/p/…) that anyone can open, no account needed — send it to a customer or a supplier. New link invalidates the old one, Turn the link off stops it immediately. The link uses the address you are using the app at (like a project link): in the cloud your instance address, on your own installation perhaps an address inside the company network — the recipient must be able to reach it.
  • People — access like sharing a project: pick named members of the organization from the list, or switch on Whole team. They find the page in the Shared with me tab (they never get the link or the token). The assistant can do it too: "share it with Jane".
  • Download as HTML — one file you can open anywhere.
  • The page comes with visuals out of the box: stat cards, progress bars, a Gantt chart or timeline by deadlines, a traffic light, icons on the sections — all drawn from the project facts (no external images).
  • Changes go through the assistant: "add a slide with the risks", "make it dark" — it rewrites the page and the previous version stays restorable. If writing fails, the panel offers Try again.
  • By hand you can only rename a page; the content is written by the assistant. The chat shows a card (Page "…" — open) that takes you to it.

A connected AI agent (Claude Code, Codex…) can upload a page too, through MCP (create_page, share_page…) — there the agent writes the whole HTML itself.

Security: the page runs in an isolated sandbox (a foreign origin in the browser) — it cannot reach your sign-in, your data or the instance API, even when a third-party agent wrote it. Public links are marked noindex for search engines. A page is one HTML file without attachments (images only linked or embedded); at most 300 kB and 100 pages per user (Cloud Lite 20; the instance administrator changes this with the KB_PAGES_* variables, KB_PAGES=0 turns the feature off). Pages are part of Download all my data. Writing a page is a longer model call — it counts as three messages in the hourly AI limit and the credits follow the real tokens.

App and organization settings ​

The assistant also handles what you would otherwise look for in the user menu (the person icon, top right) and in Organization settings. Just say it: "Switch me to English", "Turn off deadline e-mails", "Lock the alignment to compact", "Invite jane@company.com as a manager", "Set the weekly turn quota to 200". It first reads how things are set now ("What are my settings?" lists everything at once) and then:

  • Personal settings: light/dark theme, map readability, alignment lock, full and display name and turning notifications on (in-app / e-mail, per type or all at once) change right away, with Revert on the card. Language, the e-mail mode (right away / daily digest / none), turning notification e-mails off and turning all notifications off are confirmed on a card first — an instruction hidden in a step description of a shared map or in an attachment cannot switch your language or silence deadline e-mails without your consent; after confirmation the card has Revert too. A reverted setting is never re-applied by itself and the card shows Reverted. Theme, readability and the simplified view apply to this device only — just like when you switch them yourself.
  • The simplified view switches through a confirmation card (it changes the whole screen); the assistant stays at hand there and the Full app button in the header leads back.
  • Organization settings always go through a Yes, do it card and only with the permissions you have: inviting a member (administrators and managers; a role above member only an administrator), roles and the AI agents manager / org structure manager flags, deputy, organization name and purpose, AI settings on your own server (provider, address, model, transcription, images), the weekly AI quota (in turns) and the administrators' reserve, the default skin of the instance, billing details and a membership order by bank transfer (in the cloud), the AI agent registry (without the secret) and the org structure (positions, holder, deputy). A member does not get these tools at all; when they ask, the assistant says an administrator can do it.
  • An invitation without e-mail (instance without SMTP): the temporary password is shown once on the card with a Copy button — hand it over in person. It is not stored in the conversation; after a page reload it is gone.
Personal settings change right away — the card keeps a Revert link.
Organization settings always go through a confirmation card — and only with your permissions.
- **Project sharing**: *"Add Petr to the Kitchen project"*, *"Share #12 with Jane, view only"*, *"Remove Karel's access to Kitchen"*, *"Give the whole team edit access to Kitchen"*, *"Who sees the Kitchen project?"*. Same rules as the **Share** dialog: the owner or a co-manager (*Edit* permission) shares by name, team access is set by the owner only; the permissions are *view*, *collaborate* (the default when you do not say) and *edit*. A recipient without an account gets an invitation and sees the project after registering with the same address. The assistant never creates an account while sharing: on an instance with closed registration invite them to the team first (*"Invite Petr to the team"*, a confirmation card — only if you may invite), otherwise the administrators get a request in the bell.
Sharing a project: the card shows who, which permission and the current access.

Projects, rules and the rest ​

Since v0.72 the assistant also covers the rest of the app's menus — always through a confirmation card:

  • Steps and projects: "Delete the step Choose a supplier" (with its sub-steps), "Archive the Kitchen project", "Restore Kitchen from the archive", "Rename the project to Kitchen 2027", "Delete the Kitchen project" (it offers archiving first). Only the owner archives or deletes; editors can rename.
  • Public link: "Turn on the public link of Kitchen" — owner only; the link is in the Share dialog.
  • Rules and templates: "Change the rule Done → notify so that…", "Delete the rule …", "Save it as a template", "Delete the template …". When changing a rule the assistant sends the whole new shape (as the dialog does), not a single field.
  • Reminders on steps: "What reminders do I have?", "Remove the reminder on the hardware step" — the deadline stays.
  • A different deadline on someone else's work: "Ask to move the deadline of Order hardware to 20 Oct, waiting for the supplier" — the assigner gets the request and the deadline changes once they agree (in your own project the assistant changes it directly). The assigner can decline it through the assistant; accepting means setting the new deadline.
  • Notifications and reports: "Mark all notifications as read", "Report a bug: …" or "Send the developers an idea: …" (reporting needs e-mail on the instance, otherwise the assistant says so).
  • Comments, timer, moving steps, ideas, documents: "Comment on the hardware step that it arrives on Monday", "Start the timer on the hardware step" / "Stop the timer" / "Is my timer running?", "Move Buy wood under Preparation" (with its sub-steps), "Edit the idea Glue: 2 tubes", "Bring back the previous version of the inquiry" / "Delete the document Inquiry".
Archiving, renaming or deleting a project — always through a card; only the owner deletes.

What has no tool ​

Some things the assistant deliberately does not change — it only tells you where to find them:

  • the password → user menu → My account (the e-mail is only shown there, it cannot be changed; your full name and display name the assistant changes for you);
  • API keys → API keys there;
  • the AI provider token → Organization settings → AI section;
  • deleting an account and resetting a colleague's password → Organization settings → members table;
  • the secret of an AI agent → AI agent registry; company logo → Organization settings.

Passwords and secrets would otherwise pass through the conversation and stay in its history; deleting people cannot be undone. Nobody changes their own role through the assistant — another administrator does that.

When the assistant has no tool for something, it says so right away in its first reply and points to where it is in the app — it does not ask clarifying questions first that it could not act on.

Morning briefing ​

The first time you open the app on a given day (or after more than eight hours) the assistant offers "Shall we do the morning briefing?" Choose Yes, let us go or Not today. You can start a briefing any time with the Morning briefing chip.

The briefing opens by asking whether you have anything on paper or on your mind: paste a photo of your notes (Ctrl+V, drag and drop, or the paperclip / camera on the phone) and write down all ideas and tasks — the assistant sorts them, recommends what to make of them (a new project, into a project, the buffer) and writes after your consent (you confirm with a card). Nothing to add? Tap Nothing to add, go on. Then it goes through what is where, suggests where to start and asks a few click questions. At the end it asks whether to save the briefing notes — after Yes you find them in Documents under Summaries.

Evening planning ​

The Evening planning box (in an empty conversation below the morning-briefing offer, button Let's go) is for emptying your head: paste a photo of today's notes and write down whatever is still on your mind. The assistant saves nothing right away — it first sorts the items and recommends what to make of them, e.g. "a new project "Party" from 6 items, 2 into the idea buffer for later, they fit nothing". Answer Yes, do it that way, I want it differently (it asks group by group) or Everything into the buffer; creating the project and the writes are then confirmed with cards. Evening planning does not deal with today's or tomorrow's tasks — that is the morning briefing's job. It never pops up on its own and sends no reminders. At the end it offers to save the notes to Documents.

Evening planning has its own box below the morning-briefing offer — only you start it.
Box detail: one button, nothing to configure.
The assistant sorts and recommends first; it creates and writes only after your Yes.
Recommendation detail: what goes into the project, what into the buffer, and three choices.

Relax ​

When you need to switch off for a while, the assistant starts a calm scene: an aquarium, a beehive or an ant colony. Start it with the fish button 🐟 in the panel header (a second click brings the map back), the word relax in the chat, or the 🐟 Relax chip in an empty conversation. The word "relax" never reaches the model — it costs nothing. The scene opens next to the chat instead of the map, so you can keep writing to the assistant; the toolbar button switches it to full screen (and back), next to it is the browser's full screen. Pick scenes with the column of buttons at the right edge.

The countdown is up to you. Without one the fish swim until you stop them. A click on 5 · 15 · 30 min (in the scene toolbar and next to "Relax running" in the chat) starts a countdown, +5 extends it; relax 15 in the chat starts it right away. When it ends the scene chimes softly, dims and offers Back to work, 5 more min or Keep relaxing.

Lock. The lock icon in the toolbar: enter a 4-digit PIN (twice), the scene stretches to full screen and stays locked even after a page reload; unlock with the same PIN (after five wrong attempts a 30-second pause). It is a curtain for passers-by, not security — when leaving the computer, lock the system too (Win+L). Don't reuse your card or phone PIN.

The assistant offers a break itself: when you write about being tired or overwhelmed (it asks Shall we take a break? with 5 / 15 minutes / no countdown / not now and starts nothing without your consent), after 50 minutes of continuous work (at most once an hour), and as a 🐟 Relax 5 min chip after a bigger turn. In the morning, the first time you open the assistant that day, the aquarium starts by itself (once a day; not in the simplified view).

On a phone the scene takes the top 40 % of the screen with the chat below, in landscape they sit side by side like on a desktop; the same in the simplified view. In the English UI the scenes are in English too. The scenes load only while a break runs and do not work offline.

Relax next to the chat: the scene replaces the map, the assistant keeps working.
On a phone the scene sits above the chat; turn to landscape and it moves next to it.

AI helpers ​

In an empty conversation, below Evening planning, there is an AI helpers block — ready-made procedures you start with one tap: Sort my notes, New project with AI, After the meeting, Weekly review, Meeting prep and — for administrators and managers — Team meeting. On a computer it is a grid, on a phone a collapsed row AI helpers (N); tap to expand it and the app remembers that.

Sort my notes ​

Like Evening planning, just without "today" — any time during the day. The assistant asks you to paste a photo of your notes, record a voice note or write your ideas down (depending on what your instance supports). It then sorts the items and recommends what to make of them: a new project, into an ongoing project, or keep them in the buffer. It writes only after your Yes, do it that way and your confirmation on the cards.

The whole buffer: the Idea stash panel has a Sort with AI button from two ideas on. The assistant goes through the ideas already in the buffer and suggests "a new project "Workshop web" (3 ideas), 1 into the Workshop project, delete 2, keep 2" — it suggests deleting obvious test or nonsense entries. Answer Yes, do it that way, I want it differently – keep asking or Keep everything in the buffer.

Clearing the buffer: you can also tell the assistant "delete these two from the buffer" or "clear the whole buffer". It deletes nothing by itself: it shows a card Delete N ideas from the idea buffer — cannot be undone listing every idea that will go, and deletes only after your Yes, do it.

New project with AI ​

The Or let AI draft the project… link in the New project dialog (or the button in the AI helpers block) opens the assistant. What you filled in the dialog — the goal, emoji, colour and client — is not lost: the goal goes to the assistant, the emoji, colour and client go to the new map (the model does not see them).

  1. From the AI helpers block (without a goal) it starts with a form, like the former advisor: What is the project goal? (examples to click, or your own text) and How detailed should the plan be? The form shows up right away, without waiting for the AI. Instead of filling it in you can add material: text, a .txt / .md file (its content goes into the input box), a PDF, a photo or a voice note.

  2. It asks 3 follow-up questions about the goal, each with suggested answers. When you came from the New project dialog (with a goal), a fourth one is added: How detailed should the plan be?

    choicewhat you get
    Brief – 5–7 points5–7 main steps without sub-steps
    Detailed – 3 areas with 2–3 steps each (recommended)3 areas, 2–3 steps in each
    In-depth – 3 levels, 18–25 steps3 areas → 2–3 steps → 1–2 sub-steps

  3. It proposes the whole tree and shows it on a card. When the proposal does not match the chosen level of detail, the app returns it to the assistant for a fix before you see it (at most twice, then the card goes through).

  4. Yes, do it creates the project; after No it asks what to change (a different level of detail, edit the steps, discard).

This procedure replaced the former Suggest with AI and Map from text dialogs.

After the meeting ​

Right after a meeting: the app looks into your calendar (when a meeting has already started today, it mentions it — "How did the meeting "Workshop sync" go?") and at once, without waiting for the AI, asks you to paste a photo of the notes, record a voice note or write down what was agreed and who was there (I will write the notes · Nothing to write down). With the notes the assistant proposes tasks per project — "task → Jana, by 3 Oct", possibly a new project and what is unclear — and asks Do it this way? After Yes, do it that way it creates the tasks (you confirm the cards) and prepares an e-mail to the participants "who – what – by when" to copy.

When it assigns a task to a colleague who cannot see the project yet, the card says so up front — "(this gives jana@company.com access to the project)": the assignment shares the project with them so they can see the task.

Weekly review ​

A look back at the week over your own work only: what you finished in the last 7 days, what is overdue, what is stuck (not moving for long, blocking other steps, or assigned by you to someone else and overdue) and what is coming within 7 days. The app puts the overview in Done / Stuck / Next week together from your data at once, without waiting for the AI, and asks what you want to handle next week (tasks to click) and what to do with what is stuck (Break it down · Write to the owner · Leave it). The assistant then works with your answers. It plans at most 7 days ahead; it changes a deadline only on your explicit wish and the card shows the old and the new deadline.

Meeting prep ​

The app asks at once, without waiting for the AI, Which meeting are we preparing? — to click: the project you are in (first), the nearest events from your calendar, other projects and With a person – I will type the name (then it offers the people from the team to click). With your answer the assistant goes through what moved in the project over 14 days (finished, started, new steps, deadline and owner changes), for a meeting with a person their open work — only from projects you can see — and prepares materials to copy: Agenda, Open points, Stuck, What moved, To decide. Then it offers Put the meeting in the calendar (on a card) or Write an e-mail with the agenda.

Team meeting (for administrators and managers) ​

Only administrators and managers (the manager role) see it; the assistant checks the role at the start, in every step and when a card is confirmed. The app shows the team's work at once, without waiting for the AI — who has the most work, what is on fire, what is stuck — and asks What shall we do about it? After Suggest handovers the assistant proposes at most three handovers "task (project): Jana → Petr". After Yes, hand over you confirm them on cards (only the assignee changes, plus the deadline when you say so) and get a short message for the team.

Privacy: the team meeting works only with team and shared projects. Private projects — yours or anyone else's — are not in it, not even in the totals; in it the assistant does not use your personal memory, does not see which map you are in and has no tools to read private projects.

From My day straight into the assistant ​

  • Below the morning encouragement in My day there is a Discuss in the morning briefing link.
  • An overdue task has a Stuck? button (on a phone just a lifebuoy icon) — the assistant works through exactly that task with you.

Smaller AI helpers outside the panel ​

  • The wand on a node in the map editor: Substeps, Milestones, Metrics (KPI), Risks, Improve this node — one click, and the last change can be undone.
  • On any node's task list: a proposal of concrete subgoals — the message Subgoals added — N new goals were added.; the last change can be undone.
  • In the project dashboard: an AI summary of the project — a paragraph for your status report.
  • In My day: a short morning encouragement (⟳ regenerates it); without AI, proverbs rotate instead.

Ask how things are done ​

The assistant knows these guides too. Ask "How do I switch to the dark theme?", "Where do I find the idea stash?" or "How do I invite a colleague?" and it looks the answer up in the documentation (the card says Looked at: the user guide), answers briefly with the exact button names and offers More in the guide: links to the matching pages of this site under the reply. It answers only from what the guides say — it will not invent a feature the app does not have. Questions about your projects ("How is the Kitchen going?") still go to your maps, not the guides.

When the AI is down ​

An AI outage is not a problem with your data: the AI buttons in the map and in My day hide and come back by themselves once the AI responds again. Maps, tasks and sharing keep running; AI is an add-on, never a dependency.

The assistant itself does not disappear. The Assistant button stays in the header (with an orange dot) and the panel shows a "The assistant is down right now" banner (self-hosted: "Your AI isn't responding"). Conversation history, Documents, PDF (merge, split) and Relax keep working: type "relax" or tap 🐟. Messages and voice notes wait until the AI is back. A message you've typed stays in the box and the banner just blinks. Once the AI responds again, the assistant comes back on its own, or right away with Try again.

Self-hosted without any AI, the assistant is visible too, with a "No AI connected yet" invitation. Until then it offers Relax and Documents. Administrators get a Connect AI button that takes them to Organization settings → AI features (see below).

Memory ​

The assistant keeps short notes about you — preferences, style, context. You can tell it directly: "Remember that I write short e-mails." It confirms with an I will remember: line.

To see what it remembers, open Documents (the folder icon in the assistant header): the first, pinned item is What the assistant remembers — or use Open memory under the I will remember line. Unlike documents, the assistant reads this with every message:

  • What the assistant remembers about you — edit the text and Save, or wipe it with Clear all. Next time it gets exactly what you leave there.
  • Project notes — what it wrote down about individual projects (who decides, what is awaited, agreements). Each note can be edited, or removed with Delete. Whole e-mails and summaries do not belong here — they live in Documents.
Memory is visible and editable — next time the assistant gets exactly what you leave here.

Conversation history ​

  • New conversation (the plus icon) starts with a clean slate; the previous one stays saved.
  • Previous conversations (the clock icon) lists your recent conversations; click one to return to it, use the bin icon to delete it.
  • Conversations are stored on the server, so they survive a page reload and a language switch. You only see your own.

AI credits (for administrators) ​

Administrators find an AI credits section in Organization settings: this week's assistant usage for the whole organization and per person, split into Administrators and Other members, plus the previous weeks — in turns. 1 turn ≈ one ordinary message to the assistant (average of real usage; a long message with a photo or attachment costs more, a short question less). Counted from the assistant chat tokens; work of local models is free; the older AI features in the map are not counted yet.

Under Weekly quota you set Turns per week (0 = no own limit) and Administrators' reserve (%) (default 30 %). Administrators may use the whole quota; the reserve is the part other members cannot use up — other members are capped at quota minus reserve. 100 % = other members get nothing, 0 % = no reserve, everyone draws from one pool. The quota runs from Monday 00:00 UTC to Sunday and does not carry over. When a member hits the others' cap, or the whole organization quota is used up, the assistant replies that the weekly quota is used up and resets on Monday. The check runs before a turn and usage is recorded after it, so the last turn may overshoot the quota slightly.

If the operator of the instance has set a cap, you will see "The operator set a cap of N credits per week. Your own value can only lower it, not raise it."

A quota of 60 turns per week with a 30 % administrators’ reserve: administrators may use up to 60 (18 are theirs for sure), other members share at most 42.

Limits ​

  • PDF: the correction is an overlay (the original text stays in the file, the font may differ); scans without a text layer and PDFs with editing restrictions cannot be corrected; at most 60 pages / 40 000 characters of text per attachment.

  • In the simplified view (lite) the Assistant button in the header opens it full-screen — the same conversation as on the computer (close it with the arrow down). In the full app on a phone it is the bar at the bottom. When you ask it to switch to lite, it does so only after you confirm a card that says how to get back.

  • Hourly cap. One person may take at most as many turns per hour as KB_AI_MAX_PER_HOUR allows (default 60). Above it you see "You have used up the hourly AI cap. Try again in a while."

  • An image = 3 turns. A turn with an image costs more (transcription + reply), so it uses three turns of the cap (KB_AI_IMG_VAHA). An invalid image or disabled image reading does not use the cap.

  • A voice note = 2 turns (transcription + reply, KB_AI_HLAS_VAHA), at most 5 minutes (KB_CHAT_HLAS_MAX_S). An invalid recording does not use the cap.

  • The weekly turn quota, if an administrator or the operator has set one (see above).

  • A message is at most 8,000 characters — text added from a .txt / .md file too; a longer file is cut and the panel says so.

  • The assistant has no access to anything you cannot access; it reads and changes rules only in maps where you are an editor.

In the cloud and self-hosted ​

What it costs: in the cloud AI is included in the plan, with a weekly quota of turns per instance. Self-hosted, you bring your own model — a local Ollama, which costs nothing but needs a decent graphics card, or a key to any OpenAI-compatible service (OpenAI, OpenRouter, Groq, Mistral… or your own vLLM / LM Studio) — there you pay per request and need no hardware. Without a model configured the AI features simply stay hidden. The assistant stays visible with an invitation to connect an AI.

  • In the cloud the operator configures the assistant. An organization administrator can watch and limit usage in AI credits; a cap set by the operator can only be lowered, not raised. The administrator can also connect their own model in the Own model or API key section — on a PC with a graphics card at home or at work, or at a provider of their choice; turns on the own model do not count towards the quota. Steps and how to expose a home PC safely: Own model in the cloud.

  • Self-hosted, the assistant turns on as soon as the instance has a chat model:

    1. a dedicated model just for the assistant: KB_CHAT_PROVIDER, KB_CHAT_URL, KB_CHAT_MODEL, KB_CHAT_TOKEN;
    2. when empty, the daily-summary model (KB_SUMMARY_*);
    3. and when that is missing too, the general AI settings (KB_AI_* or Organization settings → AI features).

    The provider must be ollama or openai (an OpenAI-compatible API) and the model must support tool calling — e.g. gemma4, qwen3 or gpt-oss via Ollama, or gpt-4o-mini on an OpenAI-compatible service. With the api or custom provider the panel does not show.

    Images additionally need a model that can see them: KB_VISION_PROVIDER, KB_VISION_URL, KB_VISION_MODEL (e.g. gemma4 via Ollama, or an OpenAI-compatible API with image input), optionally a fallback KB_VISION_ZALOHA_*. Without variables it also works in Organization settings → AI features: with Ollama or OpenAI tick The assistant reads images (optionally a different model for images, e.g. gpt-4o-mini, gpt-4.1-mini, gemma4) and save — the app immediately tries a built-in image with text and enables images only when the model actually reads it (Test an image repeats the check). Without image reading the panel does not offer a photo (the paperclip takes PDFs only) and the morning briefing and evening planning ask for ideas as text — a model that cannot see the image would invent the transcript.

    Voice notes need speech-to-text: either an openai service that offers it (OpenAI has whisper-1 and gpt-4o-mini-transcribe; the assistant checks its model list), or your own transcriber in the OpenAI shape — KB_TRANSCRIBE_PROVIDER=openai, KB_TRANSCRIBE_URL with the full …/v1/audio/transcriptions address (speaches, whisper.cpp server, faster-whisper-server) and KB_TRANSCRIBE_MODEL. Without speech-to-text the microphone does not show. KB_TRANSCRIBE_PROVIDER=none switches voice notes off.

    Reasoning models (o-series, gpt-5*, e.g. GPT-5.6 Luna) get a token reserve for reasoning and reasoning_effort: low — adjust with KB_AI_REASONING_BUDGET and KB_AI_REASONING_EFFORT.

    A weekly turn cap for the whole instance is set with KB_AI_KVOTA_TYDEN.

All variables with their defaults are in Environment variables; the general AI setup is described in Configuration.

Next ​

fair-code — self-hosting and internal use are free, reselling as a hosted service is not.

killBottleneck