Chatting with an Agent
Open an agent from Home and start chatting. This page covers text chat; for voice see Voice Chat.
The layout
Top → bottom:
- Agent greeting — at the top of a fresh conversation before any messages are sent. Shows the agent's icon, name, and a prompt such as "How can I help you today?" It is replaced by the message stream once you send the first message.
- Message stream — your messages on the right, the agent's on the left.
- Tool call cards (when the agent uses a tool) — collapsible, show what tool ran and what it returned.
- Missing Connectors banner (when relevant) — shows above the input if the agent expects a connector you haven't authorized.
- Input area — a textarea plus icon buttons for attachments, tools, conversation history, and the primary action (Send / Voice).
At the start of each new conversation, a "New conversation started" pill appears at the top of the message stream as a visual separator.
Message formatting
Agent responses are rendered as rich text — markdown (headers, bold, lists, tables, emojis) displays fully in the chat.
Send a message
- Type your message.
- Press Enter to send. Shift+Enter for a newline.
Note: On first load, the input area is briefly disabled while the connection is established. It activates automatically — no action needed.
While the agent processes your message, a status label appears before the reply streams in. Common labels include Thinking…, Wondering…, and Considering the best next step… — these are normal and indicate the model is working. The reply then streams in, word by word.
The input area auto-grows from roughly two lines up to a maximum height — long messages stay readable while you type.
Switching to voice
There's no separate Text / Audio toggle. The right-most button next to the message input is context-sensitive:
- When the input has text, it's Send.
- When the input is empty and the agent supports voice, it changes to a Voice mode button (waveform icon) — click it to start a voice session.
To leave voice mode, click End voice mode (the X in the voice control bar). Both modes share the same conversation, so you can mix text and voice turns in one chat. See Voice Chat.
Tool calls
When the agent uses a tool, a small card appears in the stream:
Executing tool:
<tool_name>
Click to expand and see the tool's input arguments and result in a raw text format. The card auto-collapses after about 3 seconds — expand it again whenever you want to inspect it.
Most tool calls happen automatically. Two cases require something from you:
- Local tool configuration. Some agents use tools that run on your side (a device or a local MCP server). The first time one of those tools is needed, the chat opens a configuration dialog where you fill in arguments or secrets before the call can proceed.
- Manual tool responses. When an agent is set up to ask the user to provide a tool's response, a dialog opens with a JSON editor — type or paste the response and submit it back to the agent.
Conversation History modal
Click the clock icon (bottom-left of the input area) to open Conversation History. It shows:
- A search field to filter past conversations by keyword.
- A list of past conversations for this agent, each with its title and timestamp. Auto-created conversations are titled New Conversation until renamed.
- Click any entry to load that conversation into the current view.
Connecting a service mid-chat
If the agent expects a connector you haven't authorized yet, two things can happen:
- Before a new conversation starts, the dashboard may open a Missing Connectors dialog asking you to connect (or proceed without).
- During a chat, a Missing Connectors banner shows above the input. Click through to authorize.
You can also connect or reconnect services from the chat's Tools modal (the wrench icon next to the input). See Connectors.
Tools modal
Click the wrench icon to open the Tools modal. Each connector row shows:
- The connector name and type badges (Remote, OAuth, or API key).
- A Connected button (green) or Reconnect button (red) to manage authorization.
- Expand the row with the ▼ chevron to see Enabled Tools — the individual functions that connector exposes to the agent.
Stopping a reply
There isn't a separate "Stop" button while the agent is streaming. The input stays active, so you can type and send a new message at any time — the agent will start over with your latest input.
Errors and notices
| Notice | What it means |
|---|---|
| Daily quota exceeded | You've hit your usage limit for the day. The chat input is disabled until the quota resets. See Usage and Quota. |
| Connectors not connected. Some tools may not work. | The agent has tools that need your OAuth — go to Connectors and connect. |
| Connecting… state at the top | The dashboard is reconnecting the WebSocket. Hold on; messages will resume. |
| Red Reconnect button on a tool | A connector token expired. Click to reauthorize. |
Tips
- Be specific. "Summarize this in two sentences" beats "summarize briefly."
- Stay in one conversation when continuing a topic. The agent has full memory of the current thread.
- Start a new conversation for unrelated topics — open the agent again from Home or its URL; there is no separate new-conversation button on the chat page.
- If a reply looks wrong, ask follow-ups. Agents improve with more context.