Skip to main content
CollabaseAI app icon CollabaseAI is the assistant that sits inside Collabase itself. It answers questions about the work you already have in the platform, and it can act on that work: creating pages and whiteboards, opening and changing work items, writing comments. It answers in the language of your interface, whichever language you ask in. It only ever sees what you can see. Every question it answers is scoped to your own permissions, so a space, project or page you cannot open is a space, project or page it cannot read for you either.
On screen, CollabaseAI calls tasks work items and sprints cycles. Both are the same things documented under Projects and Sprints.

Opening the assistant

CollabaseAI sits in a panel docked at the right edge of the screen, below the top bar. On a screen at least 1,024 pixels wide, the panel squeezes the page beside it and does not cover it. On a narrower screen it lies over the page. Closing the panel does not end the conversation. An answer that is still coming in continues, and the chat is there when you open the panel again. If AI chats is not there, an administrator has not enabled AI on this instance yet, or you have not been given the Collabase AI app right. See AI Configuration. On the Home page, CollabaseAI also appears inline. See Your Home page and favorites.

Using the panel

Resizing

On a wide screen, drag the left edge of the panel to make it narrower or wider, between 420 and 720 pixels. The width is remembered on your device and applies the next time you open the panel.

Switch to an agent

The control at the top left of the panel names who answers. Select it to choose CollabaseAI or one of the agents you may see. A search field filters the agents by name and description. An agent chat looks like a CollabaseAI chat, with the picture and the name of the agent, its own chat history and its own conversation starters. The agent acts with your permissions. Your choice is remembered on your device. CollabaseAI keeps running in the background while you talk to an agent.

Starters

An empty panel shows a row of starting points that drifts sideways. What it offers depends on what you have open. Under the row, five categories help you word a question: Search, Create, Summarize, Analyze and Plan. Select a category to see three openers, then choose one. It is written into the input and not sent, so you finish the sentence before you send it.

The context pill

When a work item, page or Registry object is on screen, a small pill in the footer of the input names it. That is what CollabaseAI is reading, so you can ask “what changed here?” without naming the item.

The model

The footer of the input shows the model that answers, with the provider it runs at. Collabase connects one model to CollabaseAI, so the list holds that one entry. Agent chats do not show it. They always use the model of the instance.

What you can ask for

CollabaseAI works in two directions: it looks things up, and it changes things.

Finding and reading

Work items, pages, cycles, discussions, Registry objects and test runs, across every space you can access, including their history.

Creating and changing

Create work items, pages, whiteboards and Registry objects, change fields, move an item to another status or another project, and write comments. Each one waits for your approval.

Looking back

A digest of what moved while you were away, and the review numbers for a finished cycle.

Working out where you stand

What is assigned to you, what is due, what a cycle delivered against what it committed to, and whether a release milestone is ready.
Every skill it has, with an example of how to ask for it, is listed in the skill reference.

Tools of connected servers

An administrator can connect MCP servers, for example one that searches this documentation. CollabaseAI and every agent can use their tools in any chat. A tool that the server marks as read-only runs on its own. Every other tool asks for your approval first, and the card shows the server, the tool and the exact input.

Before it changes anything

CollabaseAI never writes on its own. Creating a work item or a page, changing a field, moving an item, posting a comment: each one arrives in the chat as a card naming exactly what it would do, with Approve and Reject. Nothing happens until you approve, and rejecting changes nothing. This cannot be switched off. CollabaseAI reads pages, work items and comments written by other people, and a sentence inside one of them can be phrased as an instruction to it. The approval card is what stands between such a sentence and a change made under your name. Approve a card once and the action runs once, so a dropped connection cannot produce the same work item twice. A written change then behaves exactly as if you had made it yourself: version history, notifications, mentions and the activity feed all record it under your name.

When it cannot find something

If a search comes back empty, it says so and names what it looked for. It does not guess. If you name a space or a project it has not seen, it lists the ones you do have rather than inventing a match.

Conversation history

Chats are saved as you go. Select the title of the chat at the top of the panel to see your earlier conversations, each with a short title taken from your opening question. Click one to reopen it exactly as it was, including the result cards. Delete any chat from the same list. The list shows the hint Kept for 30 days.
Saved CollabaseAI chats are kept for 30 days and then deleted for good. This is a fixed window on every Collabase instance. Chats with an agent follow other rules: you keep up to 100 of them per agent until you delete them.

Privacy

Chat transcripts are unusually revealing. They hold what you were unsure about, what you were planning, and quotes from whatever you had open. Collabase treats them accordingly.

Limits worth knowing

  • A single question scans at most 100 spaces and 200 projects. On very large instances, narrow the question to a space or project by name.
  • A registry search without a named schema or type looks through at most 50 object types, and resolves at most 10 relations per object to their names.
  • Lists come back in pages of 10, up to 25 at a time. Ask for more, or narrow the search, if the item you want is not there.
  • History and discussion are read 20 entries at a time, newest first.
  • The “what happened” digest looks back at most 90 days and returns at most 25 entries per category.
  • A status move only offers the statuses the item’s workflow actually allows from where it stands. If the one you want is not on the list, the workflow does not permit that step from here.

Next steps

Skill reference

Every skill CollabaseAI has, with an example of how to ask for it.

AI Configuration

For administrators: connecting a provider and enabling semantic search.
Last modified on October 27, 2026