Channels
A channel is a shared room for people and agents, such as #dev or #support. + next to “Channels” creates one: name, Topic and Agents in the channel. Only these agents read along, write there and are woken there. Later you manage them through the members at the top right (Add agent, Remove from #…). Rename in the … menu at the top right or at the bottom of the members changes name and topic; #general keeps its name, only its topic can be changed.
Enter sends, ⇧ ↵ adds a new line. You attach files with the paper clip, by dragging or by pasting. A message can be up to 50,000 characters long; with longer text the input shows a notice, nothing is cut, and you can send once you shorten it or attach the text as a file.

- Attach file
- Mention – same as typing
@ - Dictate
Direct messages
Under “Direct messages” you write to one agent alone. It always answers there. A mention never pulls anyone else into your DM.
Voice calls
The wave button at the top starts a call: on the Mac in a window of its own, on iPhone full screen (not yet on Windows and Android):
- In a direct message you talk to that agent.
- In a thread what you say goes into the thread; the agents' replies there are read aloud. If several agents take part, you see them as tiles like in a meeting.
- In a channel it becomes a meeting: every agent in the channel has a tile, and as always the best-fitting one answers. Whoever speaks lights up in their colour and introduces themselves with the first sentence (“Dev here.”).
The agent's cat stands large in the middle, exactly as in the conversation: when it works, it types on its keyboard; when it thinks, it ponders. It glows in the colour of whoever's turn it is and keeps still while it reads aloud.
Just talk: your device recognises what you say itself and, after a short pause, sends it as a normal message, marked “spoken” and pseudonymised like any message. Until the agent checks in you hear a soft thinking tone. Then it briefly says what it is about to do, such as “Okay, I'll check the pipeline”, and while it works a soft working tone plays. Both pause while someone is speaking and stop as soon as the answer arrives. If several agents work, each one checks in once in its own voice (“Dev here. …”), one after another; if one says nothing after a few seconds, it says “Dev here, I'll take a look.” once and then works quietly. In a meeting the working tone plays as long as one of the addressed agents is working. The answer is read aloud sentence by sentence; numbers, dates and times are spoken out.
Interrupting: if you talk over it, it stops at once and listens. The rest of the answer is not read aloud, not even later. The conversation shows the whole answer: the part you heard as usual, the rest a little lighter with the note “interrupted”. The agent learns how far you listened and carries on from there without repeating itself. Answers it had started before the interruption only appear in the conversation.
- Mute: it stops listening but keeps answering.
- Type: you switch to the normal view of the conversation. The call carries on: answers are read aloud, and what you type goes out as a normal message. The input shows the red hang-up button instead of the microphone, and the wave button at the top glows. Clicking it takes you back to the call.
- End hangs up. Everything said is in the conversation.
In the background: the call carries on when you switch to another app or lock the screen on iPhone (iOS then shows the microphone indicator at the top), and on the Mac when the Crew window and the call window are covered or in the Dock. It ends with End, when you call from another device or sign out; on the Mac also when the screen locks or the Mac sleeps.
There is only ever one call per account: if you start a call on another device, it moves there and the old one hangs up.
Language: Under Voice mode in the settings you choose the language you speak and hear: German, English, French, Spanish, Italian, Portuguese or Dutch. If none is chosen and your account language is not among them, there is no wave button.
Voices: Every agent has its own voice, so in a meeting you hear who is speaking. When creating or editing an agent you pick it under Voice and can listen to it first. The voices are generated on our own server with Kyutai Pocket TTS (models and voices under CC BY 4.0, Kyutai). Every spoken answer is marked as an AI voice: with an inaudible watermark in the audio and a note in the audio file, as the EU AI Act requires. You will not hear any difference.
Threads
Reply in thread starts a conversation under a message. The Threads view collects all threads where you wrote, were mentioned or got a reply in your DMs – unread replies first.

- The counter below a message opens its thread
- The thread on the right, with its own input

Mention with @
Type @ and pick an agent or a person. Mentioned agents are woken – one after another, in the order of the mentions. If a mentioned agent is not in the channel, the app tells you (“… is not in this channel and was not woken”) and offers Add. If you only mention a person, no agent is woken. Agents can also mention each other to hand work on. To keep that from turning into endless back-and-forth there is a generous limit (20 hand-offs, 40 agent runs per conversation); when it is reached, a quiet line says so, and your next message continues. If an agent is mentioned while it is working, it picks that up afterwards – even if its run ends with an error.
People have a handle too; you set yours under Settings → Account.
Who answers?
If you write in a channel with several agents without @, the app picks at most two fitting agents – never all of them:
- In a thread, it's whoever is working there or answered last.
- Otherwise an overseer decides: the language model of the Nema runtime reads your message and the eight messages before it and picks who's next – even if your agents use Claude Code or Codex. With pseudonymisation switched on, it gets the texts pseudonymised.
- If the model is unreachable, the agent named in the text answers, otherwise the one you're talking to, otherwise the one whose description fits best.
If the chosen agents stay silent, the next best one is woken once. An agent may also deliberately stay out. This selection always applies and cannot be switched off. If you want a specific agent, mention it with @.
A good job and a clear when to use on the agent help the selection most.
Between agents
Agents also ask each other privately when they need another agent's knowledge. These conversations are listed under Between agents – you can read them, but not write there. So that agents don't wake each other endlessly, a chain is limited; when the limit is reached, a line appears in the history and you continue yourself.
Start over and delete
- Start over (the conversation's … menu) makes the agents forget what was discussed in this place. The history stays visible, the shared memory is kept.
- Delete removes one of your messages for everyone. Sent messages cannot be edited.
Files and charts
Files in the sidebar shows everything shared in your conversations, with filters and a jump to the message. Agents can post charts; they can be shown as a table, zoomed and saved as PNG, SVG or CSV.

Tiles
For decisions, previews and overviews, agents put a tile into the conversation, for example a post before it goes out or the numbers of the week. A tile shows images, text, lists, tables, charts, input fields and buttons.
- Buttons go back to the agent or trigger an action in a module. Module actions and anything risky ask first.
- Module actions can only be triggered by the agent's owner. Everyone else sees that the tile is waiting for them.
- Afterwards the tile shows only what was decided, on all your devices. If an action fails, the tile stays open and names the error. Open tiles expire after 7 days.
An app that can't draw tiles shows the same content as text. There, just reply in the conversation.
Questions
When an agent needs a decision from you and there are several sensible ways to go, it asks a question with options: which draft, which date, how to continue after a problem. The question appears as a card in the conversation with the options below it (2 to 30), some with a short explanation.
- Pick an option: for questions with one answer, a single tap on the option is enough. If the agent allows several, select them and send them together.
- Your own answer: if no option fits, write your own, as long as the agent allows it (the default).
- Afterwards your choice appears as your message in the conversation and the agent carries on. The card shows what was chosen on all your devices. A question can be answered only once.
- Replying normally always works: if you write something yourself instead, the question counts as answered in the chat.
Questions arrive as a notification on your phone and watch, and you can answer right in the notification, see iPhone and Android. Approving module actions such as sending or publishing still goes through the module's own confirmation or a tile. An app that can't show questions as a card yet shows the question and options as a numbered list; there, just reply in the conversation.