KyoubeAI 1.9.0

Kyoube apps can now hand work to your agents. A person presses a button in an app, the right agent gets a task, and the app shows the agent's reply and the data it updated.

What's in 1.9.0

Task templates. An app declares the tasks it can send: which agent, a title, instructions with the row's and the person's values filled in, and the fields the agent may write back. A person confirms these when publishing a version that adds or changes a template, or that would send it to a different agent. An agent can't publish that version or roll back to it. A published version keeps the agents it was confirmed with, so renaming agents doesn't redirect anything.

Started by people, as themselves. Only people start or cancel tasks, and the task is created in their name. The core's own assignment rules apply, so someone who may not assign a protected agent is refused inside an app too. An agent is only asked to write Data the person can write.

Follow it to the end. The app watches the task as it goes from waiting to working to done (or needs review, blocked or cancelled). It shows the agent's reply and re-reads the rows the agent changed. It can open the task page or cancel the task. If an agent is paused, or the core won't wake it (for example when it is over budget), the start fails with a clear reason, so no task is left stranded.

Kept in bounds. App data reaches the agent in a fenced block marked as data, with size limits. Limits apply per app: 10 templates, 10 starts a minute per person and 25 open tasks. Every start and cancel is audited without row values. Company Settings → Data access lists each template, its agent and whether that agent can write Data.

Update

From 1.8, run ./update.sh. It backs up first, and ./update.sh --rollback returns to 1.8. The update adds one empty table and changes nothing until a person publishes an app version with task templates.

How app tasks work, and what they don't cover, is in docs/app-tasks.md. The full list of changes is in CHANGELOG.md.

KyoubeAI 1.8.0

Kyoube apps and agents can now work with the other services your company uses, such as payments, CRMs, shipping and weather, through connections an admin sets up. API keys never reach the browser.

What's in 1.8.0

Connections. An instance admin adds a connection under Settings → Plugins → Kyoube Data & Apps, with the company selected: a name, the service's https address, how it signs in (a Bearer token, an API-key header, or Basic) and a company secret holding the key. Each connection is read-only or read-write. Company Settings → Data access shows whether each one is ready, which apps use it, and which agents may call it.

Apps that call out. An app names the connections it uses and calls them with kyoube.connections.call. It works as the person using it: groups decide who can open the app, and changes at the other end need that person's write access. When a new app version adds a connection, or turns a read-only one read-write, a person publishes it after seeing exactly which services it can reach. An agent can't.

Agents that call out. An owner or admin can let an agent use a connection, read-only or read-write. With the guardrail on, an agent's write waits for a person's OK. The Kyoube Apps skill teaches agents how to find connections, use them, and what to ask you to set up.

Kept where you pointed it. A call can only reach addresses under the connection's base address. Tricks that try to step outside it are refused, and every call is recorded without its contents.

Update

From 1.7, run ./update.sh. It backs up first, and ./update.sh --rollback returns to 1.7. The update adds one empty table and changes nothing until an admin adds a connection.

How connections work, and what they don't cover, is in docs/connections.md. The full list of changes is in CHANGELOG.md.

KyoubeAI 1.7.0

KyoubeAI can now keep teams apart on one instance: user groups decide who may use which agents and apps.

What's in 1.7.0

User groups. Owners and admins create groups under Company Settings → Groups. A group holds people, agents, apps and, if you like, a Data level. An agent or app that is in no group stays open to everyone, so nothing changes until you create your first group. Owners and admins are never restricted.

Agents only their group can use. Outside a group, people can't give the group's agents work or chat with them. They can still see the agents, their tasks and their runs. Changes reach agents within about a minute.

Apps only their group can open. A group's apps disappear from everyone else's Apps page, and a link to one says "You don't have access to this app. Ask a company admin." The app's tables are not hidden: they stay on the Data page at each person's own level.

A Data level per group. A group can set read, write or schema, and the highest level a person has wins. Groups can lower an operator to read. They can't raise a viewer, since viewers can't make changes anywhere in KyoubeAI. The Data access page now shows each person's level and where it comes from.

Part of the licence. Creating and changing groups needs a KyoubeAI licence. If the licence lapses, your groups keep working and you can still delete them.

Update

From 1.6, run ./update.sh. It backs up first, and ./update.sh --rollback returns to 1.6. The update adds empty tables and changes nothing until someone creates a group.

For agent rules and groups to work in a company, the instance's board key user must be a member of that company (any role). How groups work, and what they don't cover, is in docs/groups.md. The full list of changes is in CHANGELOG.md.

KyoubeAI 1.6.0

KyoubeAI now runs on the latest Paperclip core, and chats with pi and Hermes agents read the way they should.

What's in 1.6.0

pi agents say things once. A pi agent's replies and thinking used to appear four or five times in a row in task chat. Each one now shows once, and still streams in while the agent works.

Hermes agents are readable. Hermes agents ran in a noisy terminal mode even though the agent form showed Quiet output switched on, so the chat filled with the agent's own instructions, text cut at 80 characters, and boxes. They now run quietly unless you switch Quiet output off, and a reply shows as one message with its paragraphs, lists and tables intact. Runs from before the update still show the old output when you open their details.

The latest core. KyoubeAI moves to Paperclip 2026.1005.0. Agent harnesses no longer stop to ask permission for their own tools, as KyoubeAI agents mostly did already. KyoubeAI's checks on what agents may do still apply. Mentioning an agent in a comment no longer starts it working: assign the task or ask for a review instead. Agent Chat, a page for ongoing conversations with an agent, is there to try under Settings → Experimental.

Agent working rules keep your edits. When you edit an agent's instructions while KyoubeAI is adding its working rules, your edit is kept, and the rules are added to it within a minute.

Update

From 1.5, run ./update.sh. It backs up first, and ./update.sh --rollback returns to 1.5. The new core runs 14 database migrations on the first start, so give it a few minutes before it answers.

What changes for your team is in docs/upgrading.md. The full list of changes is in CHANGELOG.md.

KyoubeAI 1.5.0

KyoubeAI now installs as an app on your phone and computer, and tells you when something needs you, even when it isn't open.

What's in 1.5.0

Notifications when you're needed. KyoubeAI sends a push notification when an agent asks you something, when an approval is waiting for a decision, and when a task you created or are assigned is done or blocked. You can also choose to hear when an agent's run fails (owners and admins) and when someone comments on your tasks. Tap a notification and KyoubeAI opens on that task or approval.

Install it like an app. On iPhone and iPad, open KyoubeAI in Safari and add it to your Home Screen. On Android and in Chrome or Edge on a computer, install it from the browser. A card at the top of Home walks you through it and turns notifications on for that device. The new Notifications page lets you send a test, pick what you hear about, and remove devices you no longer use.

Private by design. Your KyoubeAI sends each notification itself, encrypted, straight to the phone's or browser's own push service. There is no KyoubeAI server in between, and the push service can't read what it carries. Notifications show only a task or approval name, never comment text or app data, because lock screens are visible to others.

Better on a small screen. At most two pop-up messages show at a time, above the tab bar. Kyoube Apps open full width with a compact header, and the Data page shows your tables first, then one table at a time. Agents now build apps that work on a phone as well as a laptop.

Update

From 1.4, run ./update.sh. It backs up first, and ./update.sh --rollback returns to 1.4.

Notifications need KyoubeAI to be reached at an https:// address (phones don't allow notifications from plain http://). If yours already is, there is nothing to set up: each person turns notifications on from the card on Home. If not, docs/mobile.md explains the options.

The full list of changes is in CHANGELOG.md.

KyoubeAI 1.4.1

KyoubeAI 1.4.1 lets your agents change their own and each other's setup when you allow it, and fixes the team links on the dashboard.

What's in 1.4.1

Agents no longer ask for a permission you can't give. An agent that tried to update its own instructions, or another agent's instructions, skills or settings, was refused and asked you to grant "agents:configure". There was nowhere in KyoubeAI to grant it. Now there are two ways:

• Your top agent has it from the start. The agent that reports to nobody, usually the first one you hired, can now change your agents' setup, as a CEO agent can in the core. If more than one agent reports to nobody, none of them gets it automatically.
• A switch on every agent's profile. Open an agent from the Team list and use Can change agents under Permissions to give it this permission or take it away. Only company owners and admins can flip it. Switching your top agent off keeps it off.

Dashboard team links work. Clicking an agent under "Your team right now" on Home showed "Organization not found". It now opens the agent's profile.

Update

From 1.4.0, run ./update.sh. It backs up first, and ./update.sh --rollback returns to 1.4.0.

With the agent working rules on (the default), your top agent gets the new permission within a minute of the update. You can switch it off on the agent's profile.

The full list of changes is in CHANGELOG.md.

KyoubeAI 1.4.0

KyoubeAI can now use Jev, a decision model from TypeSafe AI, to make quick judgments for your agents, your data and your apps, and to check risky agent actions before they run.

What's in 1.4.0

Jev decides, it doesn't write. You give it some information (a support ticket, a row of data, a description of what an agent is about to do) and a few closed questions. It answers each one in a fraction of a second: which of these options fits, how urgent is this on a scale, is this statement true. Every answer comes with a confidence score, so KyoubeAI knows when to go ahead and when to ask a person. Jev only ever picks from the answers you allow, so its output is always usable, and it is cheap to run at volume.

KyoubeAI 1.4.0 puts Jev, or any model that works the same way, to work in four places:

• Your agents can ask it. Agents use it for quick judgments at scale, such as sorting tickets into queues, rating urgency or spotting refund requests. Anything it isn't sure about goes to the person who started the task.
• Data tables can fill themselves in. Add an AI column to a table, write the question, and Jev fills it in for every row and keeps it up to date as rows change. Answers it isn't sure about wait for a person to accept or change them, and a person's edit always wins.
• Kyoube Apps can make decisions. An app can ask Jev about a record, for example to route a ticket to the right team, with a review lane for uncertain answers. Before such an app goes live, a person sees exactly what it will send and approves it.
• A safety check on risky agent actions. When you switch it on, Jev checks whether an action fits the agent's task before the agent drops a table, deletes many rows or publishes an app. Anything doubtful waits for a person's go-ahead, and only a person can release it.

You stay in control. Each company uses its own key, for TypeSafe directly or through OpenRouter or Vercel's AI Gateway, and turns each use on separately, with a daily limit. Nothing is sent anywhere until you do.

Update

From 1.3, run ./update.sh. It backs up first, and ./update.sh --rollback returns to 1.3.

Everything new starts switched off. To try it, add your provider key in the Kyoube Data & Apps plugin settings, then switch on the uses you want under Company Settings → Data access. The update lets the plugin reach the provider you choose, but nothing is sent until you set it up.

The details are in docs/decisions.md, and the full list of changes is in CHANGELOG.md.

KyoubeAI 1.3.0

KyoubeAI now has a free limit of 5 users per instance. A licence key from KyoubeAI raises it.

What's in 1.3.0

• Free for up to 5 users. Once an instance has 5 users, the next new account is refused, with a message saying an admin can add a licence key. Agents never count, and everyone who already has an account keeps working.
• Licence keys raise the limit. An instance admin pastes the key under Settings → Plugins → KyoubeAI Licence, or runs kyoube license set <key>. The page then shows "Licensed to …" with the number of users and the expiry date.
• Checked offline. Keys are verified on the instance itself. Nothing is sent to KyoubeAI, and no internet connection is needed.
• Expiry is gentle. Admins get a reminder from 30 days before a key expires. After it expires the free limit applies again, and nobody is locked out.
• Free a seat with kyoube users list and kyoube users remove <email>.
• kyoube doctor reports the licence and whether the limit is enforced. It can now print WARN lines, which don't fail it.

Update

From 1.2, run ./update.sh. It backs up first, and ./update.sh --rollback returns to 1.2.

If your instance already has more than 5 users: everyone keeps working after the update, but no new user can be added until you apply a licence key. kyoube doctor shows a WARN line until then.

The details are in docs/licensing.md, and the full list of changes is in CHANGELOG.md.

KyoubeAI 1.2.0

Agents now finish the task you give them, and approvals come to you. You no longer get a chain of handoffs and approval requests passed between agents.

What's in 1.2.0

• Agents finish their own tasks. An agent does the work end to end. It does not pass the task to other agents or ask them to review it. It hands work on only when you ask for that in the task, or when its new ## Handoffs section says so.
• Approvals go to the person who started the task. Only a person can answer an agent's approval cards.
• Work moves up to a manager agent only with your decision. An agent that can't do a task hands it back to you with an "Escalate" or "Keep" choice.
• Managers can still split a big task across their own team.
• Kept in force automatically. kyoube agent-rules re-applies the rules once a minute, and kyoube doctor reports them. The core is unchanged.
• Can be turned off. Set KYOUBE_AGENT_RULES=off and run kyoube agent-rules off to remove them. The details are in docs/agent-rules.md.

Update

From 1.1, run ./update.sh. It backs up first, and ./update.sh --rollback returns to 1.1.

From 1.0, run git fetch --tags, then git checkout v1.2.0, then ./update.sh.

The full list of changes is in CHANGELOG.md.

KyoubeAI 1.1.0

Install and update in one command each, and harnesses you choose and install yourself.

What's in 1.1.0

• One-command install. ./install.sh takes a fresh git clone to a running, claimed instance with its plugins. It is safe to run again.
• One-command update. ./update.sh updates to the newest release after taking a backup. An interrupted update continues when you run it again. ./update.sh --rollback returns to the release you came from.
• A smaller image. The image no longer includes pi, Hermes Agent or a pinned Claude Code, which makes it about 2.5 GB smaller. Install the harnesses you want from the Terminal with kyoube harness install <name>. They keep their logins across restarts and updates.
• Packages persist. The node user has passwordless sudo. Packages you install with apt are reinstalled at every start.
• Claude subscription connections now last a year instead of about 8 hours, via kyoube connect claude. Connections made before 1.1 must be connected again once.
• Fixes:
  • Test now works on an unsaved harness switch.
  • Harnesses installed from the Terminal are found by agents as well as by Terminal shells.
  • Public instances allow subscription sign-in and local MCP tools.

Update

From 1.0:

1. Run git fetch --tags.
2. Run git checkout v1.1.0.
3. Run ./update.sh.

The full list of changes is in CHANGELOG.md.