Your ClawBox has two chat surfaces on the desktop, and they talk to the same assistant:
  • the chat panel, which opens when you click the mascot and docks beside your windows;
  • the Chat app, a full window from the desktop icon.
Everything on this page is on both editions unless it says otherwise.
The chat panel open over the desktop with an empty transcript: a small crab above the line Say something!, the message box below it, and the provider and reasoning pills in the row beneath.

The chat panel on a conversation nobody has said anything in yet

The chat open over the desktop on a short exchange: a greeting, then a question asking what the assistant can help with on this ClawBox, and its reply saying it can manage files on the box, use the browser, run shell commands and call its skills. The composer below carries an attach button, a plus button and a model pill reading Gemma 4.

The same chat with something in it — this box is answering from Gemma 4, on the box itself

​
The pills: provider, model and reasoning

The chat panel carries up to three pills in the row beneath the message box. A change applies to the running agent. There is no restart between you and the next message, and open conversations are not dropped.
On ClawBox AI there is no model pill — the chat names the provider and there is nothing to choose. The reasoning control is there and starts at Medium.
A reasoning change is a per-conversation override. It does not rewrite the box’s default, which lives in Settings → Providers.
The chat panel with the provider pill open above the message box: the connected provider ClawBox AI with a tick beside it, the on-device model Gemma 4 E2B offered below it, and the reasoning pill reading Medium alongside.

The provider pill open on a box running ClawBox AI, with the on-device model to switch to

​
Slash commands

Type / in the composer and a menu appears with the commands your assistant actually publishes — the built-in ones plus anything your skills and plugins added. Each row shows the command, its arguments if it takes any, and the agent’s own one-line description. Keep typing to filter; pick a row to insert it. The list is read from the agent itself, so it is never out of date with what your box can actually run, and it differs between editions and between boxes with different skills installed.

​
When the assistant needs you

Three kinds of card can appear in the transcript. All three mean the assistant has stopped and is waiting for you.
The assistant asks you to choose. Each option carries its own short description — usually the consequence of picking it — and there is a field for an answer of your own. One Send answer submits the whole card. There is no skip: to decline, leave it, and the assistant carries on when its own timer runs out.
The assistant is asking to do something that needs your permission. Allow once, Always allow or Deny. Nothing happens until you answer, and if nobody answers in time the window closes and it does not go ahead.
The assistant wrote a message and wants to send it from your account. You read it first and nothing goes out until you approve — this card is what Ask me before sending in Settings → Email produces, and it is on by default. You can send them all, send one, or delete them without sending. See Messaging channels.

​
When a message fails

A failed turn says what actually went wrong: the provider’s own reason, not a generic “the run failed”. Two common cases have their own wording:
  • A fallback model answered. Your first-choice model refused or errored and the next model in the chain replied instead. The reply says which model wrote it.
  • An allowance ran out. On ClawBox AI, the chat says whether it was the weekly chat allowance, the 5-hour burst or the memory-indexing allowance, and when it frees up. Your usage is in Settings → Providers.

​
Voice

Voice in — in the chat panel, press the microphone to record. The recording is turned into text and sent as a message. The Chat app has no microphone. Recording needs a secure connection, so on a plain http:// address the chat points you at this box’s Remote Access tunnel instead of failing silently. Voice out — Settings → Voice → Spoken replies, off by default. With it on, a reply gets a player on its bubble: press play, scrub through it, or download the audio. A question you asked out loud is answered out loud. Where your box answers messaging channels, a voice note there gets a spoken reply too.
Speaking on Telegram, WhatsApp and Discord is an OpenClaw-edition feature. On the Hermes edition the box still answers out loud in the chat; replies on those channels stay in text.

​
Other things in the composer

  • Attach a picture or take a photo, and ask about it.
  • Generate a picture from a description.
  • Several chats at once, in the chat panel — tabs across the top, each its own conversation. New session starts a fresh one in the same tab.
  • Files the assistant produced appear as cards you can download, with their real name and size.

​
Where the conversation runs

The pills are per-conversation; the box’s default provider and model are in Settings → Providers, and that is what a new conversation starts on. When a sign-in has to restart the agent — connecting a subscription, for instance — the chats you already have open are re-pointed at the new default once the agent answers again, so a conversation cannot keep running on the old model while the pills name the new one. If that could not be done, the box tells you rather than leaving it to be discovered.