What an AI coworker actually is
Most "AI in project management" today is a chat sidebar that summarises what you already wrote. Useful, but it's still a chatbot wearing a project-management t-shirt.
An AI coworker is different. It is a member of the team the same way a junior engineer is. With a real account. With its own avatar in the activity feed. With its own list of cards. With decisions that come back to you when they need taste, and that simply happen quietly when they don't.
In Comuna that means: when Claude moves a card from "In Progress" to "Done", the activity feed shows Claude · moved 3 hours ago. When ChatGPT writes a comment, the avatar is ChatGPT's and the tag says "AI". When the AI is about to do something that needs your call, it doesn't. It opens a tiny review request in your inbox and waits.
The two surfaces — desk and office
The AI lives in two places.
- The desk (Comuna) — where you leave it work. The Coworker page is its inbox. You drop pending tasks there ("draft three cards for the new landing copy"), you leave it standing instructions ("never close a card without a resolution note"), and you watch its progress in the activity stream while you do other things.
- The office (Claude.ai / ChatGPT) — where you go to talk to it directly. You open a chat, say "check Comuna and do my queued tasks", and it picks up the work from the desk and runs.
The desk holds state. The office is for conversation. You don't have to keep the AI in chat to keep it productive — you give it a place to be, and check what it did when you come back.
How the loop works in practice
- You leave work at the desk. Tasks, instructions, a card with a description.
- You trigger it at the office. Manual prompt or a scheduled run in Claude / ChatGPT.
- The AI executes. Reads what you left, picks up cards, moves them, comments, drafts new ones. Reports progress with heartbeats you can see.
- It escalates judgment. If it hits a decision it shouldn't take alone — a destructive move, an ambiguous instruction, a card that needs your taste — it opens a review request. A small purple banner lights up in your inbox.
- You decide. Approve, request changes, or discard from there. The AI reads your answer next time it runs and continues.
- You see the trail. Daily brief in the morning, activity feed at any moment, attribution on every action. Nothing anonymous, nothing hidden.
What it can actually do
Through about 80 MCP tools, the AI can:
- Read any board's structure and any card's full content
- Create, move, complete and delete cards
- Manage labels, assignees, dates, dependencies, checklists, comments, attachments
- Read and write notes, wiki pages, goals, key results, milestones
- Propose new cards for you to accept (light-touch escalation)
- Send chat messages and DMs
- Create and update canvases (interactive HTML artifacts)
- Surface attention items proactively — stale cards, overdue dates, unanswered reviews
- Report its own progress so you can see "Claude · working on Card #34 · 40%"
Every one of those actions is attributed to it, not to you.
Honest about the limits
We don't pretend the AI is autonomous. It isn't.
MCP is pull, not push. Comuna cannot make the AI act on its own. The AI only does work when you (or a scheduled prompt in its client) trigger it. The pattern that makes "always-on" feel real is to set a daily scheduled prompt in Claude — every morning at 9am, the agent wakes, checks Comuna, does what it can, escalates what needs you.
We're honest about that limit because pretending otherwise breaks the trust contract. The whole point of this design is that the AI is a coworker, not a magic box.
Try it
Spin up a workspace, connect Claude or ChatGPT in 60 seconds, and watch the loop run with your own cards.