AI Agent

A chat assistant that lives inside your workspace. Ask it about your work in plain English, or tell it to do something — create a work item, move it to In Progress, add it to a sprint — and it will do it, after showing you exactly what it is about to change and waiting for your approval.

Plan: Pro and above

What you can do

  • Ask questions about your real data. "What's overdue across my projects?", "Which urgent items are unassigned?", "What do we know about the login bug?" It reads your workspace and answers with real references like APP-42.
  • Ask how the product works. "How do cycles work?", "What's the difference between a module and an epic?" No data is touched for questions like these.
  • Tell it to make changes. Create work items, projects, cycles, modules and labels; update state, priority, title, description, assignees, labels and due dates; add comments; add items to a cycle or module; make an item a sub-item of another; delete an item.
  • Approve or reject every change. Nothing is written until you say yes.
  • Keep your chats. Every conversation is saved, listed in a sidebar, and can be reopened or deleted.
  • Get triage suggestions while you write. The agent also powers the AI suggestions on the work-item form (see below).

Getting started

Turn on AI

AI features need an AI provider configured once, for the whole instance, by an admin:

  1. Open the admin console (God Mode) → Artificial Intelligence.
  2. Choose a provider — OpenAI, Anthropic (Claude), Google Gemini, Groq or OpenRouter.
  3. Pick a model and paste the API key. The key is encrypted at rest.
  4. Save. Every AI feature in every workspace now uses that provider and model.

Until this is done, the agent replies with a message telling you AI is not configured.

Open the agent

  1. Look at the icon rail on the far left of the app — the same strip that holds Projects and Settings.
  2. Click the AI Agent (robot) icon.

📷 Screenshot: the left app rail with the AI Agent icon highlighted, and the empty chat state next to it.

The empty state offers four starter prompts you can click straight away:

  • Find work items about onboarding
  • What's overdue across my projects?
  • Create a task to fix the login page
  • How do cycles work?

Type in the box at the bottom. Enter sends; Shift+Enter adds a new line. While the agent is working you can press the stop button to cut it off.

Watching it work

The reply streams in as it's generated. Above the text you'll see small action chips showing what the agent is doing and what it did — "Searching all work items", "Listing members", then "Created APP-42" with a green tick, or a red cross if something failed.

Read actions (listing, searching, fetching details, project stats) run on their own. They never change anything.

Approving changes

The moment the agent wants to write something, it stops and asks.

  1. An amber Approval required card appears with a plain-English summary of the change — for example Create issue "Fix login redirect" in APP, assign to Priya, in cycle "Sprint 12".
  2. Choose Accept or Reject on each card, or use Approve all / Reject all in the bar at the bottom.
  3. As soon as every pending action has a decision, the agent carries on: it runs the approved ones, is told which were declined, and then writes its reply.

📷 Screenshot: an approval card in the chat with Accept and Reject buttons, and the Approve all / Reject all bar below.

Anything the agent creates or updates goes through the normal product API — so activity history, notifications and webhooks all fire exactly as if you had done it by hand.

AI triage on the work-item form

When you create or edit a work item, an AI triage suggestions panel appears beside the form once the title is a few words long.

It suggests:

  • a priority,
  • up to three labels,
  • one assignee,
  • and a one-line explanation of why.

The suggestions are constrained to labels and members that actually exist in that project, and they're grounded in how the project's recent items were really triaged — so they follow your team's conventions rather than generic guesses.

The panel is advisory. It does not fill the fields for you: set the values yourself in the fields above it.

📷 Screenshot: the work-item create form with the AI triage suggestions panel showing a priority chip, label pills and an assignee.

Tips

  • A question is not a command. "Can you create a project?" gets an answer; "Create a project called Atlas" gets an approval card. If you want action, use an imperative.
  • Ask for one thing per message when it matters. The agent will chain steps, but a clear single instruction is easier to review before approving.
  • Use names, not IDs. "Assign APP-12 to Priya", "label it backend", "put it in Sprint 12" — names are resolved against the project's real members, labels and cycles. If a name doesn't match anything, the agent tells you instead of inventing someone.
  • Relative dates work. "Due Friday" or "end of the month" is converted to a real date.
  • Reopen an old chat from the history sidebar to continue with its full context. Collapse the sidebar with the panel button in the header if you want more room.

Limits & good to know

  • It only ever acts as you. Writes are made with your permissions — it cannot see or change anything you couldn't see or change yourself.
  • Every write needs approval. There is no "just do it" mode in the chat.
  • Deletes are possible but deliberate. The agent will only propose deleting a work item if you clearly ask, and you still have to approve it.
  • It works one turn at a time, with a cap on how many tool steps it can take in a single reply. Very large multi-step requests are better broken up.
  • Only the recent part of a long conversation is sent to the model, so extremely long chats may lose early context. Start a new chat for a new topic.
  • Lists are capped. "List everything" style questions return a bounded set (tens of items, not thousands) — ask a narrower question for precision.
  • It can be wrong. It is told never to invent references, dates or numbers, and results are checked against real records — but always sanity-check an answer before acting on it.
  • Provider rate limits surface as errors. If your AI provider throttles the request, you'll see a "rate limit reached" message; wait and retry.