A Kanban board you talk to

Things people actually say

• —Move the billing migration to In Review.
• —New card: chase the design feedback, due Friday, high priority.
• —Comment on TAR-142. The client approved the copy.
• —What is blocked, and who am I waiting on?
• —Open the Tarsen board.

Every line maps to a tool the model has. It can read the board, create and update cards, move them, assign them, label them, comment, run checklists and lists, archive work, and open any screen in the app.

How a sentence becomes a change

1. 01

   Press ⌘J

   Your browser asks for the microphone once. The OpenAI key stays on our server. The browser is handed a token that expires in ten minutes and does nothing else.

2. 02

   Talk normally

   Tarsen decides you have finished by what you said, not by how long you paused. You can think mid-sentence. You can also interrupt the reply.

3. 03

   The server checks the work

   The model never touches the database. It asks for an action, and the same permission checks the board uses run again against your account.

4. 04

   The board moves

   The change appears live for everyone watching. The card's activity records the sentence that caused it.

The model is OpenAI's gpt-realtime-2.1 and the voice is called cedar. Audio runs straight from your browser to OpenAI, because that is the only way to keep a reply fast enough to interrupt.

What it stops to ask

Speech can be misheard. Text cannot. So two actions wait for a spoken yes: deleting a card, and creating a project. Tarsen reads the name back to you, and writes nothing until you agree.

Voice can open your billing page or your member list. It cannot change either one. There is no tool for that.

Voice acts as you. It sees the projects you can see and nothing else. A private project stays private, and the activity log names you as the actor.

What happens to what you say

• —No audio is stored. There is nowhere in the product that could hold a recording.
• —The text of a conversation is kept for ninety days, then deleted.
• —Only the person who spoke it can read it. There is no admin view of someone else's session.
• —Each change carries the one sentence that caused it. Anyone who can see the card can read that line.

That trade is deliberate. A colleague can find out why a card moved without reading your whole conversation.

What it costs to talk

Minutes reset on the 1st. Realtime audio is billed to us by the minute, so the allowance is a number we can stand behind rather than a promise of unlimited talking.

What you need, and where it stops

• —A browser and a microphone. There is no app to install and no headset to buy.
• —A connection. A voice session is live, so it does not work offline.
• —An admin switches voice on for the workspace. Setting up a new workspace offers it.

Voice is more than dictation, though it can type into a field you have focused when you ask it to. It is the same assistant that runs the board, listening. If you would rather type, the keyboard does everything voice does.