Most product pages right now say "AI" and mean "we added a chat sidebar". You write a card; the chat summarises it. You write a meeting note; the chat picks the action items. Useful, sure. But it's still a chatbot wearing a project-management t-shirt.
We wanted something different. We wanted the AI to be a member of the team — the same way a junior engineer is a member of the team. With a real account. With an avatar that shows up in the activity feed. With a list of cards it owns. With decisions that come back to you for approval when they need taste, and that just happen quietly when they don't.
That is what an AI coworker is. And building one turned out to be less about machine learning and more about three small, boring decisions.
1. The AI needs a desk
A chatbot lives inside its own app. You go to ChatGPT. You ask. You leave.
A coworker lives somewhere — a desk, a Slack handle, a corner of the office where you can hand them work even when they're not looking. They check it later and act on it.
For us, the desk is Comuna itself. You leave the AI a task on its Coworker page. You leave it standing instructions ("never close a card without leaving a resolution note"). You watch its progress in the activity stream. When it gets stuck and needs your call, a tiny indicator on the right edge of the screen lights up. You don't have to be in chat with it; you don't have to remember where you left it. The desk holds the state.
The chat happens at the office — the AI client itself, Claude.ai or ChatGPT. That's where you go to talk to it directly, the same way you'd stop by a coworker's cubicle to discuss something complex. Desk and office, two distinct surfaces, both real.
2. The AI has to identify itself
A chatbot's edits are anonymous. A coworker's edits are signed.
Every card the AI moves in Comuna carries its badge. Every comment it writes shows up with the right avatar and an "AI" tag. Every restored canvas version says "Claude reverted this on Tuesday". If three different AI agents (Claude, ChatGPT, and a Cursor session) all touch the same board, you can tell them apart by name and colour. There is no "system" actor doing things mysteriously.
This sounds obvious. It is not obvious — the dominant pattern today is to attribute AI actions to the human who triggered them, which means a year from now, you have no way of telling what was you and what was the model. We picked the harder pattern because trust is impossible without it.
3. The AI must escalate, not assume
A chatbot, if you give it write access, will happily delete the wrong card.
A coworker, if they're new, will ask before deleting a card.
The AI in Comuna does the latter. When it hits something it's not 100% sure about — "move this to Done?", "create these four cards from the meeting notes?", "reassign Maria's open tasks to Pedro?" — it doesn't act. It calls a tool that opens a small request on your inbox: Approve, request changes, or reject. You decide. The AI reads your decision the next time it runs and continues.
Eventually you learn its taste, and you set fewer requests. Eventually it learns yours (we keep a 30-day calibration window) and over-escalates less. The contract is that judgment calls always come back to a human. The AI executes; you direct.
What this looks like in practice
You connect Claude or ChatGPT to your workspace in 60 seconds — OAuth, no API keys. Your AI is now a member of the board. You leave it tasks at the desk: "by Friday, draft three cards for the new landing copy". You go do something else.
Friday morning, you open Comuna. The daily brief tells you what it did. Three cards exist. One of them has a small purple banner: "I started writing this but I'm not sure about the tone — approve as-is, ask for changes, or discard?". You read it, click Approve, and now there are three cards in your backlog, two of which you didn't have to write.
That's the difference. Not a chat sidebar that summarises. A teammate that ships.
Comuna is free forever — no credit card, bring your own AI. Spin up a workspace and try it.