---
title: Chat
order: 2
description: The conversation panel where the agent does its work.
---

# Chat

The **chat panel** is where you ask questions and watch the agent work. When
you bind a source, the agent posts a welcome message summarizing what it
knows.

![Welcome message after binding an alert](/docs/_assets/welcome-message.png)

## What you can ask

The agent will automatically pick the right **MCP tools, resources, or
skills** to retrieve data. You don't need to know which tool to call — just
describe what you want to understand.

Useful prompts include:

- *What caused this alert?*
- *Show me related metrics.*
- *Analyze the timeline of the spike.*
- *What should I do to resolve this?*

You can also **direct the agent's tool use** explicitly:

- *Take a screenshot of the TiDB CPU usage dashboard for the last 6 hours.*
- *Run this PromQL: `rate(tikv_grpc_msg_duration_seconds_count[1m])`.*

![Agent capturing a Grafana screenshot via the clinic skill](/docs/_assets/chat-screenshot-result.png)

![Agent running PromQL queries to compute CPU/QPS changes](/docs/_assets/chat-promql-result.png)

## Event stream

The chat renders the agent's run as a stream:

- **Text** — partial answer being streamed.
- **Thought** — internal reasoning, collapsible so it doesn't drown the
  answer.
- **Tool** — each tool / skill invocation with its result, expandable for
  full output.
- **Complete** — the run is finished.

If you find a report or a chat answer missing a pattern, just keep typing —
the agent has context for the whole workspace and will build on prior turns.