Skip to main content

Integrations and Widgets

An integration is the AI agent experience you publish to users. The widget is the interface users interact with.

Create an Integration​

  1. Open Integrations.
  2. Select New Integration.
  3. Enter an integration name.
  4. Select the bridge or deployment to attach.
  5. Choose the model.
  6. Add integration instructions.
  7. Add allowed domains.
  8. Enable optional features as needed.
  9. Save the integration.

Integration Instructions​

Integration instructions tell the AI agent how to behave.

Include:

  • What the AI agent should help with.
  • What it should not do.
  • How it should handle missing information.
  • When it should ask the user for clarification.
  • Any company-specific tone or policy.

Allowed Domains​

Allowed domains control where the widget can run.

Add only domains that should load the widget, such as:

https://example.com
https://app.example.com

Use staging domains for testing before production.

Optional Features​

Depending on your workspace, an integration can enable:

  • Knowledge Base.
  • Voice calling.
  • Host capabilities.
  • Canvas.
  • Correction Memory.

Each optional feature should be tested before going live.

See Correction Memory if you want the AI agent to learn from approved corrections and keep that learning useful across model changes.

Widget Display and Size​

The widget can run in compact mode or sidebar mode. The install snippet can also include optional width and height controls for desktop and mobile layouts.

Most integrations should start with the default snippet. Add custom sizing only when your website layout needs it.

For details, see Installing the Widget.

Native MCP Integration​

If you do not want to use the hosted widget, your own LLM client or agent framework can connect to the integration over MCP SSE.

For details, see Native MCP Integration.

Integration Key​

The integration key identifies the public integration used by the widget. Keep it in the widget install snippet and rotate it if it is exposed somewhere it should not be.

Going Live Checklist​

  • Correct deployment is attached.
  • Model is selected.
  • Integration instructions are complete.
  • Allowed domains are set.
  • Widget install snippet is copied from the latest integration key.
  • Widget mode and sizing are tested on desktop and mobile.
  • Double-hop auth is configured if needed.
  • Correction Memory is enabled if the integration should improve from accepted corrections.
  • Optional features are tested.