v0.9.2

Start from nothing and let the board tell you when it needs you. This release creates a project from the launcher, keeps a card out of the way until it is really ready, and notifies you the moment a run is waiting on your answer.

What changed

• Create a project from the launcher. Switching to another project opens it in its own window, the same as switching boards.
• An empty repository is a fine place to start. Setup no longer stalls on a project the repository says nothing about.
• A card stays inactive until it is finished being created, so a half-written card is never picked up.
• A notification when a run is waiting on you. "Questions waiting" fires when a card is held at landing, only when the card can actually take an answer, and delivery retries right away after a transient failure.
• Configuration → Agents is a picker column beside the agent's page, and each agent shows a localized display title.
• In-progress cards animate, so a running card reads as running at a glance.
• Draft comments are stored as editable markdown, so what you wrote can be edited.
• Planning keeps only useful memories and honors opt-outs, and the discussion clears once a plan has been handed off.
• Cloud event reads retry when they fail, instead of dropping the update.
• Windows finds the Codex CLI bundled with the Windows Desktop app on its own.
• Conflict handling keeps your changes, the resume prompt says what to do next, and the uncommitted-changes error reads clearly.
• Inbox is hidden unless you were invited to it — it is not ready yet.

Downloads

System	File	Signed	Tested
macOS (Apple Silicon)	AI4Kanban-0.9.2-arm64.dmg	no	yes
macOS (Intel)	AI4Kanban-0.9.2.dmg	no	yes
Windows (x64 + arm64)	AI4Kanban-Setup-0.9.2.exe	no	no
Linux (x64)	AI4Kanban-0.9.2.AppImage	no	no
Linux (arm64)	AI4Kanban-0.9.2-arm64.AppImage	no	no

macOS is the one we test each release. Windows and Linux are built and published untested — tell us if one doesn't open.

Opening it the first time

No build is signed yet, so every system warns once. The download page spells out every click.

• macOS — drag the app in from the .dmg, double-click it, click Done on the warning, then System Settings → Privacy & Security → Security and Open Anyway next to the blocked app — once more when it asks.
• Windows — SmartScreen says "Windows protected your PC": More info → Run anyway.
• Linux — chmod +x AI4Kanban-*.AppImage and run it.

From a terminal

npx ai4kanban@latest install

ai4kanban-ui on npm stays frozen and deprecated — the app is how the board is handed out now.

Full changelog: v0.9.1...v0.9.2

v0.9.1

Think an idea through before it becomes a card. This release turns the chat rail into a place you hold several discussions at once, hand pictures to an agent, and let a proposer read what you already shipped before it suggests what's next.

What changed

• Hold several discussions at once. Each one is listed in the rail and stays there, so a half-formed idea waits while you take another. Archiving one from the rail actually archives it.
• A discussion helper for deciding what is worth building. Talk an idea over and the helper tells you what it would cost and what it would displace, instead of writing the card straight away.
• A proposer reads your finished cards first. What it suggests next reflects what you already shipped, not just what is on the board.
• Hand an agent a picture. Drag images into the chat rail and into Discuss, paste one into Add task or Build now, and see the attachments laid out properly in the box.
• Pick the runtime for an Add task or Build now run, instead of taking the board's default for everything.
• A run retries itself when the provider blinks. A momentary provider failure no longer costs you the run.
• The board says who is holding a card, so a card stuck in someone's hands is visible at a glance.
• Every board opens in its own window. Two projects, two windows, and the notifications stay with the right one.
• Swipe back out of every view, not only card pages.
• Review and resolve a Cloud card in the browser, without going back to the machine that holds the code.
• Marketing work is a shorter loop. A new topic opens straight into its draft, the draft is verified and fixed in one capped polish pass, and the writing feedback you give is grouped into its own memory file per topic — so a correction on one topic stops leaking into the others.
• Inbox takes anything, not just market signals.
• Prune memory from the Memory panel, not from a card on the board.
• Specialist agents go by their role name, and the reviewer's switch is on the Agents pane with the rest of them.

Downloads

System	File	Signed	Tested
macOS (Apple Silicon)	AI4Kanban-0.9.1-arm64.dmg	no	yes
macOS (Intel)	AI4Kanban-0.9.1.dmg	no	yes
Windows (x64 + arm64)	AI4Kanban-Setup-0.9.1.exe	no	no
Linux (x64)	AI4Kanban-0.9.1.AppImage	no	no
Linux (arm64)	AI4Kanban-0.9.1-arm64.AppImage	no	no

macOS is the one we test each release. Windows and Linux are built and published untested — tell us if one doesn't open.

Opening it the first time

No build is signed yet, so every system warns once. The download page spells out every click.

• macOS — drag the app in from the .dmg, double-click it, click Done on the warning, then System Settings → Privacy & Security → Security and Open Anyway next to the blocked app — once more when it asks.
• Windows — SmartScreen says "Windows protected your PC": More info → Run anyway.
• Linux — chmod +x AI4Kanban-*.AppImage and run it.

From a terminal

npx ai4kanban@latest install

ai4kanban-ui on npm stays frozen and deprecated — the app is how the board is handed out now.

Full changelog: v0.9.0...v0.9.1

v0.9.0 — your board, wherever you are

Your board now lives wherever you are — put it on Cloud and open it in a browser, on a borrowed laptop or on a phone, and invite your team in. Everything that runs it is an agent you can train.

What changed

• Keep your board in Cloud and open it in a browser. A workspace we host renders the board without your machine, so a second computer or a phone opens the same cards — and the screens now lay out for a small width. Your code and agents stay where they are. Invite-only preview.
• Invite teammates into a workspace, with roles, so more than one person works the same board.
• The chat rail is a full conversation. Copy, resend or reword a message without retyping it, open the agent's steps and what the reply cost, and pick the model for one conversation instead of for the whole board.
• Create task opens as a full-screen chat, not a dialog. Talk a vague idea into a plan you can read in one screen, write it up as a task, or build it straight from the box with no card at all.
• Every flow the board runs is a named agent you train. One Agents pane replaces the Spec agents and Rules panes: give each agent a rule of your own, and memory that survives the run. Agents are grouped into roles, and the writer can call in a write agent.
• The gater and the decider are agents of their own, so what judges a card ready and what answers its open questions each take your rules and keep their own memory.
• A plan you talked through hands straight to the build. The old plan-ask step is gone: the conversation ends in a handoff the builder reads, not another question.
• The notification bell carries a dock badge, and what it raises belongs to the board you have open — a second board no longer shouts at you.
• AI review runs once per delivery, re-checks only what a rebase changed instead of the whole delivery again, and can be switched off for a board or for a single build. Each run is grouped under its delivery and says why the review started.
• Two more runtimes: Grok Build (xAI's coding agent) and Antigravity CLI — and adding one is a shorter flow than it was.
• QA effort scales to the request. A clear, small request gets a light pass; a mixed one keeps single-agent QA; a vague, broad one runs QA in parallel.
• Open a board at a path other than docs/kanban, so one repository can hold more than one board.
• Marketing work has a home on the board — repurpose a source draft per channel, and see a marketing card with its channels.
• The app reports how it is used — feature usage and failures, never code, card text, project names or file contents. Onboarding says so, the privacy page says how long a raw event is kept, and it can be turned off.
• Skip the agent step when a coding agent already works, and the landing page is rebuilt around the planning bottleneck.

Downloads

System	File	Signed	Tested
macOS (Apple Silicon)	AI4Kanban-0.9.0-arm64.dmg	no	yes
macOS (Intel)	AI4Kanban-0.9.0.dmg	no	yes
Windows (x64 + arm64)	AI4Kanban-Setup-0.9.0.exe	no	no
Linux (x64)	AI4Kanban-0.9.0.AppImage	no	no
Linux (arm64)	AI4Kanban-0.9.0-arm64.AppImage	no	no

macOS is the one we test each release. Windows and Linux are built and published untested — tell us if one doesn't open.

Opening it the first time

No build is signed yet, so every system warns once. The download page spells out every click.

• macOS — drag the app in from the .dmg, double-click it, click Done on the warning, then System Settings → Privacy & Security → Security and Open Anyway next to the blocked app — once more when it asks.
• Windows — SmartScreen says "Windows protected your PC": More info → Run anyway.
• Linux — chmod +x AI4Kanban-*.AppImage and run it.

From a terminal

npx ai4kanban@latest install

ai4kanban-ui on npm stays frozen and deprecated — the app is how the board is handed out now.

Full changelog: v0.8.1...v0.9.0

v0.9.0-rc-1

Superseded. This was the first 0.9.0 build. The shipping v0.9.0 is newer and is what the download page hands out — get that one instead.

Your board now lives wherever you are — put it on Cloud and open it in a browser, on a borrowed laptop or on a phone, and invite your team in. Everything that runs it is an agent you can train.

What changed

• Keep your board in Cloud and open it in a browser. A workspace we host renders the board without your machine, so a second computer or a phone opens the same cards — and the screens now lay out for a small width. Your code and agents stay where they are. Invite-only preview.
• Invite teammates into a workspace, with roles, so more than one person works the same board.
• The chat rail is a full conversation. Copy, resend or reword a message without retyping it, open the agent's steps and what the reply cost, and pick the model for one conversation instead of for the whole board.
• Create task opens as a full-screen chat, not a dialog. Talk a vague idea into a plan you can read in one screen, write it up as a task, or build it straight from the box with no card at all.
• Every flow the board runs is a named agent you train. One Agents pane replaces the Spec agents and Rules panes: give each agent a rule of your own, and memory that survives the run. Agents are grouped into roles, and the writer can call in a write agent.
• AI review runs once per delivery, re-checks only what a rebase changed instead of the whole delivery again, and can be switched off for a board or for a single build. Each run is grouped under its delivery and says why the review started.
• Two more runtimes: Grok Build (xAI's coding agent) and Antigravity CLI.
• QA effort scales to the request. A clear, small request gets a light pass; a mixed one keeps single-agent QA; a vague, broad one runs QA in parallel.
• Open a board at a path other than docs/kanban, so one repository can hold more than one board.
• Marketing work has a home on the board — repurpose a source draft per channel, and see a marketing card with its channels.
• The app reports how it is used — feature usage and failures, never code, card text, project names or file contents. Onboarding says so, and it can be turned off.
• Skip the agent step when a coding agent already works, and the landing page is rebuilt around the planning bottleneck.

Downloads

System	File	Signed	Tested
macOS (Apple Silicon)	AI4Kanban-0.9.0-arm64.dmg	no	yes
macOS (Intel)	AI4Kanban-0.9.0.dmg	no	yes
Windows (x64 + arm64)	AI4Kanban-Setup-0.9.0.exe	no	no
Linux (x64)	AI4Kanban-0.9.0.AppImage	no	no
Linux (arm64)	AI4Kanban-0.9.0-arm64.AppImage	no	no

macOS is the one we test each release. Windows and Linux are built and published untested — tell us if one doesn't open.

Opening it the first time

No build is signed yet, so every system warns once. The download page spells out every click.

• macOS — drag the app in from the .dmg, double-click it, click Done on the warning, then System Settings → Privacy & Security → Security and Open Anyway next to the blocked app — once more when it asks.
• Windows — SmartScreen says "Windows protected your PC": More info → Run anyway.
• Linux — chmod +x AI4Kanban-*.AppImage and run it.

From a terminal

npx ai4kanban@latest install

ai4kanban-ui on npm stays frozen and deprecated — the app is how the board is handed out now.

Full changelog: v0.8.1...v0.9.0

v0.8.1 — set the board up by talking to it

Open a folder and tell the board what you are building. It asks what it needs to know, writes the board from your answers, and puts the first cards up — no form to fill in.

What changed

• Set the board up by talking to it, not by filling in a form. Opening a project starts a conversation: it asks what the project is and how its work divides, writes config.md from what you said, and seeds three cards to start from — each one refined in the background while you read the first. Opened the wrong folder? One click undoes it and puts every file back.
• Install a new version from inside the app. A chip in the header says a newer version is out; click it and the app downloads and installs it. A version you don't want is skipped for good, and the chip goes quiet until the next one.
• Say what to build next. Ask the board what to pick up and it answers from what is on it, rather than leaving you to read every card.
• Keep typing while a reply is still coming, and stop a reply while it is being written. The chat no longer locks you out mid-answer.
• Codex is a runtime the board can run on, alongside Claude Code, opencode and ZCode — and ZCode's 0.16.5 protocol is read correctly, so token counts and tool calls show up again.
• Configuration → Runtimes says which machine each runtime runs on, so a board shared across computers tells you where its work can actually go.
• Answer a multiple-choice question in your own words instead of picking one of the options.
• A card records what the agent decided and what you overruled, in sections the scaffold writes for every new card.
• The header links to the repo, and akb refine --effort picks how hard a refinement works.

Downloads

System	File	Signed	Tested
macOS (Apple Silicon)	AI4Kanban-0.8.1-arm64.dmg	no	yes
macOS (Intel)	AI4Kanban-0.8.1.dmg	no	yes
Windows (x64 + arm64)	AI4Kanban-Setup-0.8.1.exe	no	no
Linux (x64)	AI4Kanban-0.8.1.AppImage	no	no
Linux (arm64)	AI4Kanban-0.8.1-arm64.AppImage	no	no

macOS is the one we test each release. Windows and Linux are built and published untested — tell us if one doesn't open.

The Windows installer is AI4Kanban-Setup-0.8.1.exe from this release on. Earlier releases carry AI4Kanban Setup <version>.exe and keep that name.

Opening it the first time

No build is signed yet, so every system warns once. The download page spells out every click.

• macOS — drag the app in from the .dmg, double-click it, click Done on the warning, then System Settings → Privacy & Security → Security and Open Anyway next to the blocked app — once more when it asks.
• Windows — SmartScreen says "Windows protected your PC": More info → Run anyway.
• Linux — chmod +x AI4Kanban-*.AppImage and run it.

From a terminal

npx ai4kanban@latest install

ai4kanban-ui on npm stays frozen and deprecated — the app is how the board is handed out now.

Full changelog: v0.8.0...v0.8.1

v0.8.0

Press Implement once, and the card lands itself. When it still needs you, it finds you.

Until now a card went out to a run and came back needing you to look at it. This release closes the loop: the board builds the work, reviews it against what the card asked for, corrects what review found, and lands it on your branch. What genuinely needs a person — a review to approve, a question only you can answer — leaves the machine and reaches you where you are.

What changed

• One Implement click carries the card the whole way — it builds the work, reviews it against what the card asked for, corrects what review found, lands it as one commit on your branch, and finishes the card. Nothing asks you again in between.
• Watch a delivery from the card itself — what it is doing or waiting for, the diff while it builds, and the commit that landed. Each delivery builds in a git worktree of its own, so the checkout you are sitting in is left alone.
• Approve the tree before it lands. Turn on Require diff approval before landing and no delivery reaches your branch until you have approved the exact tree it would land.
• A card that needs you now finds you away from the machine. Reviews that are ready and questions only you can answer fill a notification center in the app and arrive in Slack carrying the card, where Implement is pressed and questions answered in the message — and run when your machine is next reachable. Cloud carries them, in preview behind an invitation code you can ask for from the app.
• Configuration → Runtimes replaces the single harness setting: the board names the runtimes its work runs on, and each computer says what it runs them as.
• 简体中文. The app reads in Simplified Chinese as well as English and opens in the language your machine is set to, and the board is now written in the language you pick — cards, open questions and the lines to check included.

Also here, from 0.7.2, which never shipped a build of its own:

• Closing a release writes its changelog in plain words at the top of its record, instead of leaving only a list of cards.
• A specialist agent can carry settings of its own, not just an on or off switch. Pick them in Configuration → Agents, on the agent's own row, and every run after that uses them.
• ui-design can draw a card's layout options in ASCII instead of rendered mockup screens, so the drawing reads from a terminal too and costs far less to produce.
• Building a group task no longer burns the ids it reserves, so the numbers on your board stop jumping.

47 cards in all — the full lists are in docs/kanban/.release-summaries/0.8.0.md and docs/kanban/.release-summaries/0.7.2.md.

Downloads

System	File	Signed	Tested
macOS (Apple Silicon)	AI4Kanban-0.8.0-arm64.dmg	no	yes
macOS (Intel)	AI4Kanban-0.8.0.dmg	no	yes
Windows (x64 + arm64)	AI4Kanban Setup 0.8.0.exe	no	no
Linux (x64)	AI4Kanban-0.8.0.AppImage	no	no
Linux (arm64)	AI4Kanban-0.8.0-arm64.AppImage	no	no

macOS is the one we test each release. Windows and Linux are built and published untested — tell us if one doesn't open.

Opening it the first time

No build is signed yet, so every system warns once. The download page spells out every click.

• macOS — drag the app in from the .dmg, double-click it, click Done on the warning, then System Settings → Privacy & Security → Security and Open Anyway next to the blocked app — once more when it asks.
• Windows — SmartScreen says "Windows protected your PC": More info → Run anyway.
• Linux — chmod +x AI4Kanban-*.AppImage and run it.

From a terminal

npx ai4kanban@latest install

ai4kanban-ui on npm stays frozen and deprecated — the app is how the board is handed out now.

Full changelog: v0.7.1...v0.8.0

v0.7.1

Talk to the board, and it does the work you just agreed on.

A chat rail runs down the right of the app. On the board it is about the whole project; on a card it is about that card, and nothing of one leaks into another. It already knows the goal, the module map, the open cards and the memory — the same thing a coding agent knows after /kanban — so you never explain the project first. When the two of you settle on a change, it makes it: writes the card, rewords it, answers its open questions, moves it in or out of a release, archives it. Real building still goes to a run, and that run joins the runs panel like any other.

What changed

• Chat with the board, or with one card, in a rail that follows what is on screen. Fold it away and the reply keeps arriving.
• It does the board work it is talking about — the change lands straight away, and it says what it did. No "shall I?".
• The same conversation in a terminal: akb chat and the rail land in the same place, turn after turn.
• A failed run shows on the card it was working on — open the card and its log is already open, with a line saying the card may be part-built. Resume sits beside it.
• A card now reads in two halves — what a person has to decide on at the top, and what the agent worked out folded under one control. Search reaches into the folded half and opens it at the word.
• Run the board on ZCode, Z.ai's coding agent for GLM — the sixth harness, and the way to use a Z.AI Coding Plan.
• Planning quality, next to Daily progress — three scores read out of the board's own record: the share of card questions it settled itself, the share of its own calls you left standing, and the share of its proposals that got built. Per release, with every card behind each figure.
• Finish notes have their own field instead of squatting in an open question.

16 cards in all — the full list is in docs/kanban/.release-summaries/0.7.1.md.

Downloads

System	File	Signed	Tested
macOS (Apple Silicon)	AI4Kanban-0.7.1-arm64.dmg	no	yes
macOS (Intel)	AI4Kanban-0.7.1.dmg	no	yes
Windows (x64 + arm64)	AI4Kanban Setup 0.7.1.exe	no	no
Linux (x64)	AI4Kanban-0.7.1.AppImage	no	no
Linux (arm64)	AI4Kanban-0.7.1-arm64.AppImage	no	no

macOS is the one we test each release. Windows and Linux are built and published untested — tell us if one doesn't open.

Opening it the first time

No build is signed yet, so every system warns once. The download page spells out every click.

• macOS — drag the app in from the .dmg, double-click it, click Done on the warning, then System Settings → Privacy & Security → Security and Open Anyway next to the blocked app — once more when it asks.
• Windows — SmartScreen says "Windows protected your PC": More info → Run anyway.
• Linux — chmod +x AI4Kanban-*.AppImage and run it.

From a terminal

npx ai4kanban@latest install

ai4kanban-ui on npm stays frozen and deprecated — the app is how the board is handed out now.

Full changelog: v0.7.0...v0.7.1

v0.7.0 — a card can call a specialist

Cards can call in a specialist now, and a card that draws a screen shows you the screen.

A spec is no longer one agent's single pass. When a card needs a judgement outside the run's depth, it hands that part to an agent that only does that — ui-design draws the screen the card needs, recommend-tech-stack picks the library or the service — each in its own context, writing its part back into the card. You choose which ones are on from the board.

What changed

• Spec agents — a card calls a specialist for the part of its spec that needs one, and the specialist runs in its own context.
• Two ship in the box — ui-design for the screen a card needs, recommend-tech-stack for the library or service it should use.
• Switch them on or off in the app — the board shows every spec agent it knows about.
• See a card's mockup as a picture, not as markup to read.
• A card marks the calls that are not just common sense, so the judgement in a spec is visible.
• Configuration → Skill is now Configuration → Agent setup, in the app and in all five languages.
• A UI design reference the board's own UI work goes through.

9 cards in all — the full list is in docs/kanban/.release-summaries/0.7.0.md.

Downloads

System	File	Signed	Tested
macOS (Apple Silicon)	AI4Kanban-0.7.0-arm64.dmg	no	yes
macOS (Intel)	AI4Kanban-0.7.0.dmg	no	yes
Windows (x64 + arm64)	AI4Kanban Setup 0.7.0.exe	no	no
Linux (x64)	AI4Kanban-0.7.0.AppImage	no	no
Linux (arm64)	AI4Kanban-0.7.0-arm64.AppImage	no	no

macOS is the one we test each release. Windows and Linux are built and published untested — tell us if one doesn't open.

Opening it the first time

No build is signed yet, so every system warns once. The download page spells out every click.

• macOS — drag the app in from the .dmg, double-click it, click Done on the warning, then System Settings → Privacy & Security → Security and Open Anyway next to the blocked app — once more when it asks.
• Windows — SmartScreen says "Windows protected your PC": More info → Run anyway.
• Linux — chmod +x AI4Kanban-*.AppImage and run it.

From a terminal

npx ai4kanban@latest install

ai4kanban-ui on npm stays frozen and deprecated — the app is how the board is handed out now.

Full changelog: v0.6.1...v0.7.0

v0.6.1 — your agent, and what it remembers

The board runs on the agent you already use, and it shows you what it remembers.

Five agents drive it now — Claude Code, Codex, Cursor, OpenCode and DeepSeek Harness — picked from one screen in the app, each with its own provider, model and key. And the memory the agent writes as it works is no longer a folder you have to go read: the rail opens what shipped, the decisions that are settled, the mistakes not to repeat, and every module's own note.

What changed

• Pick your agent — Cursor and OpenCode join Claude Code and Codex, and DeepSeek Harness (dsh) makes five.
• Read the memory in the app — the board's, the project's, and each module's, from the rail.
• Install the akb command from the app — one button, no npm.
• The setup strip says when a setup run fails, instead of sitting there looking busy.
• A warmer site, and the plain-Markdown mirror of it back in line with the pages.

9 cards in all — the full list is in docs/kanban/.release-summaries/0.6.1.md.

Downloads

System	File	Signed	Tested
macOS (Apple Silicon)	AI4Kanban-0.6.1-arm64.dmg	no	yes
macOS (Intel)	AI4Kanban-0.6.1.dmg	no	yes
Windows (x64 + arm64)	AI4Kanban Setup 0.6.1.exe	no	no
Linux (x64)	AI4Kanban-0.6.1.AppImage	no	no
Linux (arm64)	AI4Kanban-0.6.1-arm64.AppImage	no	no

macOS is the one we test each release. Windows and Linux are built and published untested — tell us if one doesn't open.

Opening it the first time

No build is signed yet, so every system warns once. The download page spells out every click.

• macOS — drag the app in from the .dmg, double-click it, click Done on the warning, then System Settings → Privacy & Security → Security and Open Anyway next to the blocked app — once more when it asks.
• Windows — SmartScreen says "Windows protected your PC": More info → Run anyway.
• Linux — chmod +x AI4Kanban-*.AppImage and run it.

From a terminal

npx ai4kanban@latest install

ai4kanban-ui on npm stays frozen and deprecated — the app is how the board is handed out now.

Full changelog: v0.6.0...v0.6.1

v0.6.0 — the board app is the way in

The board app becomes the way in. Open it, point it at a project, and it asks the three things only you can answer — one to a screen, everything prefilled — then reads your codebase, maps its modules, and writes the first ten cards. No terminal, no Node, nothing to install first.

Underneath, akb is now the one way the board is run. The rules moved into the CLI, the skill shrank to a note that says the board is here, and the UI does its work through the same command an agent runs — so anything the app can do, an agent can do too.

What changed

• The app is the way in — guided first run, add the coding agent skill from a button, open and switch projects the way an IDE does.
• akb owns the board — start, watch, stop and resume a run from the CLI; the UI goes through it.
• Schedule an action on a blocked card so it runs when the blocker clears.
• A refine follows the run that touched the card — the auto-refine switch is gone.
• Find a card by typing in the rail, and swipe back and forward in the app.

23 cards in all — the full list is in docs/kanban/.release-summaries/0.6.0.md.

Downloads

System	File	Signed	Tested
macOS (Apple Silicon)	AI4Kanban-0.6.0-arm64.dmg	no	yes
macOS (Intel)	AI4Kanban-0.6.0.dmg	no	yes
Windows (x64 + arm64)	AI4Kanban Setup 0.6.0.exe	no	no
Linux (x64)	AI4Kanban-0.6.0.AppImage	no	no
Linux (arm64)	AI4Kanban-0.6.0-arm64.AppImage	no	no

macOS is the one we test each release. Windows and Linux are built and published untested — tell us if one doesn't open.

Opening it the first time

No build is signed yet, so every system warns once. The download page spells out every click.

• macOS — drag the app in from the .dmg, double-click it, click Done on the warning, then System Settings → Privacy & Security → Security and Open Anyway next to the blocked app — once more when it asks.
• Windows — SmartScreen says "Windows protected your PC": More info → Run anyway.
• Linux — chmod +x AI4Kanban-*.AppImage and run it.

From a terminal

npx ai4kanban@latest install

ai4kanban-ui on npm stays frozen and deprecated — the app is how the board is handed out now.