Hosted chat and widget
Any agent can have a chat page hosted by your Orchestrate console, and a widget that embeds the same chat in your own website. Both run the agent's live version.
Switch it on
- Open the agent and go to Connect, then End-user chat.
- Turn on Chat is on.
- Optionally turn on Allow people who are not signed in, and set how many anonymous messages per hour the agent accepts in total.
- Set how it looks: a display name, a logo, an accent colour, a welcome message, up to four starter prompts, and the language of the page.
- List the allowed origins: the websites that may embed the widget.
- Save. You need the admin role on the project.
The section then shows the page link, https://<your console>/chat/<chat id>, and the widget snippet:
<script src="https://orchestrate.example.com/widget.js"
data-chat-id="<chat id>" data-color="#2563eb" data-label="Support" async></script>
Paste the snippet into your site's pages. Only the origins you listed can frame the chat.
Who can chat
| Mode | Who | Runs as |
|---|---|---|
| Signed in (default) | Anyone with an account on this Orchestrate install. They don't need to be a member of the agent's project. | The person. Runs & traces and approval requests name them, and a connector that acts as each user uses their own connection. |
| Anonymous | Anyone with the link or on a page with the widget. | The project. Tools run with the project's connector credentials. |
An anonymous chat is rate limited twice: per visitor per minute, and per agent per hour across the whole install. A conversation is continued by its thread id, a long random value that only that visitor's browser holds.
API keys are refused on the chat. An application should use the Agent API instead: a key in a web page is exactly what the hosted chat avoids.
What people see
- The agent's answer, and the knowledge-base passages it was grounded on, with a short snippet each.
- When a tool call needs approval, the name of the tool and that the agent is waiting. The answer arrives on the page when someone approves, even after a reload.
- A thumb up or down on each answer, with an optional comment. See Online evals and feedback.
- An attach button for images, when the agent's first step uses a model that reads images. See Images.
They don't see tool arguments, tool results or error details. Those are for the people who build the agent, in Runs & traces.
For a code agent, the answer streams onto the page as the agent writes it.
Branding
A chat uses its own logo and accent colour first, then the organization's branding, then the install's. See Settings and branding.
Turning it off
Turn off Chat is on to stop the page and the widget at once. Deleting the project also switches off its chats, and restoring the project leaves them off.