How integrations work

Connect a tool from a card with numbered steps, test each action with real inputs, and control what the agent may read.

Integrations give the agent something to do beyond answering: check an order, read a subscription, file a ticket, alert your team. Each one is a card under App Integrations with the exact steps for that tool, and each creates ready-made actions: the individual things the agent can do, such as "Check order status".

How connecting works

Pick a tool under App Integrations
The catalog covers Stripe, Shopify, Slack, Zapier, HubSpot, Zendesk, Freshdesk, Zoho Desk, Salesforce, WooCommerce, Jira, Mailchimp and Discord. Anything else can be connected as a custom action.
Connect with one click or an API key
Some tools offer Connect in one click: you approve access on the tool's own consent screen and the actions are created and enabled. The others show numbered steps for creating a key or webhook in the tool, then a form to paste it into, and a checklist of the actions to add.
Test each action
Every action card has Test. It runs the real request with sample inputs and shows the status and the response exactly as the agent would see it. Testing works before you enable the action.
Enable it
Actions you add with a key start switched off. Flip the switch on the card once the test looks right. One-click connections start enabled.

Tests are real

The Test button performs the real operation. Testing "Create Zendesk ticket" creates a ticket; testing "Cancel subscription at period end" schedules a real cancellation. Use test records.

What the agent sees and does

Each enabled action becomes a tool the agent may use when a visitor's request matches the action's when to use text. You can edit that text.
The agent collects any inputs the action needs ("What's your order number?") before calling it, and never invents values. Inputs with a known format, such as digits-only order numbers, are checked before any request goes out.
The agent may only read the response fields listed on the action. Everything else in the tool's reply, including personal data, never reaches the model.
It answers in plain language and never pastes raw API output. If the tool returns "not found" it says so and asks the visitor to double-check; on other errors it offers a person.
Actions run only while the agent is answering (unresolved conversations) and never in voice calls.

Security

Keys, tokens and secrets are encrypted before they are stored and are never shown again in the dashboard; only header names are visible. Webhook URLs (Slack, Discord, Zapier) are the credential and appear on the action card, so rotate them in the tool if they leak.
Every call goes over HTTPS to the tool's own servers. Private and internal hosts are blocked, redirects are not followed, requests time out after 10 seconds, and responses are capped in size.
Visitor inputs are escaped for the exact place they land (URL, JSON, form or XML), so a visitor cannot smuggle anything into the request.
Execution is at-least-once: a rare retry can run a write twice. Do not wire refunds or payments as actions.

Limits and housekeeping

Up to 20 actions per workspace, integrations and custom actions combined.
Any workspace member can connect, edit and test integrations.
To rotate an expired key (Jira and Mailchimp keys expire yearly), delete the affected actions and add them again with the new key; the same-name guard prevents adding a duplicate.
There is no "disconnect" button: switch actions off or delete them, and revoke the key or app in the tool itself.