---
title: "Chat with the AI Agent"
id: "1944"
type: "page"
slug: "agent"
published_at: "2026-09-30T22:28:16+00:00"
modified_at: "2026-09-30T23:50:35+00:00"
url: "https://xedant.com/agents/proxy/docs/agent"
markdown_url: "https://xedant.com/agents/proxy/docs/agent.md"
excerpt: "The AI agent chat is a way to ask right in the interface: “why so…"
---

# Chat with the AI Agent

[https://xedant.com/agents/proxy/docs/agent.md](https://xedant.com/agents/proxy/docs/agent.md)

The AI agent chat is a way to ask right in the interface: “why so many errors yesterday”, “help me set up a channel”, “what does this chart show”. The answer comes not from a model directly, but from an external Xedant Agent connected to the product. The conversation happens where the data lives — no switching to another window and copying numbers around.

## How to connect it

It is enough to set two environment variables on the server:

- **`PROXYAGENT_AGENT_API_URL`** — the address of your Xedant Agent.
- **`PROXYAGENT_AGENT_API_KEY`** — its access key.

Both variables are read **once at startup**. That means after changing them the container needs a restart — saving in the interface is not enough. Until both are set, the chat counts as not connected.

## What stays on the server

The agent key never reaches the browser. It is stored on the server and inserted only into calls to the agent — from the browser window it is visible neither in replies nor in live updates. That matters when several people use the interface: the agent key stays yours.

## What it looks like

- **The chat panel** — on a wide screen it is a separate panel next to the content; on a narrow one it slides up from the bottom. One and the same panel, in two forms.
- **The chat list and the message queue** — the side panel shows which chats are working right now and which messages are waiting their turn.
- **Live updates** — the answer appears as it becomes ready, as in an ordinary chat; the “working” state, the counters and the queue refresh themselves.

Questions the agent asks along the way can be answered right in the chat — such answers are handled separately from ordinary messages, so they do not get stuck in the queue.

## What the chat panel can do

- **Chat labels** — a conversation can be given a label and sorted by topic or client. The chat list offers filtering: with labels, without labels, or by any chosen label; a chat without one is marked accordingly.
- **The task list** — above the conversation sits the list of tasks the agent keeps as it works. You see what is done, what is being done and what remains — no need to ask again.
- **Instant sending** — your message appears in the conversation immediately, without waiting for the server’s reply. If something goes wrong, it returns to the input field.
- **Back to the latest message** — in a long conversation, once you have scrolled up, a jump-to-bottom button appears.

## Sounds and voice input

A new message can be heard: the notification sound also fires for chats that are not currently open. If it gets in the way, there is a mute button next to it, and unmuting works the same way.

A message can be dictated — the voice-recording button in the input panel. How the speech is recognized depends on the settings:

- **By the browser** — the browser itself recognizes the speech, and the recording never leaves for your server. But not every browser supports it, and the browser may need an internet connection.
- **Through the connected agent** — with the other engine selected, the recorded audio goes to Xedant Agent for transcription. Without a connected agent this input does not work.

Honestly: when the browser cannot recognize speech and no agent is connected, the button will not help. And the browser needs microphone permission — without it the product suggests checking the device settings.

## How much an answer costs

The cost of every answer is visible in the conversation. Cheap answers show to the hundredth of a cent — as `$0.00xx` — and the very smallest as “under a cent”. That way you can see that a cheap model is really cheap, and not “free” because of rounding.

## Without a connection

When no agent is connected, the product works in full: models, keys, logs, analytics, secret hiding, the proxy service — everything is in place. The chat panel simply does not appear. Nothing breaks and nothing “waits for configuration”.

The bottom block of the side panel — the summary of active chats with the “Prompts” and “Chats” buttons — is always shown: open it and you will see what is missing. A warning strip appears above the buttons naming the required variables (`PROXYAGENT_AGENT_API_URL` and `PROXYAGENT_AGENT_API_KEY`); as soon as the connection works, the strip disappears.

## An honest caveat

The chat is a **separate product**. For it to work, Xedant Agent has to be installed and configured on its own; Proxy Agent merely connects to it. If you do not have the agent yet, you do not need this capability — everything else works without it.

Next: how to host the product on a server and where the data lives — in the [Hosting & Data](/agents/proxy/docs/hosting)
 section.

[← Back to the documentation index](/agents/proxy/docs)
