Manage connections
Every assistant that completed the OAuth flow is listed in Chatway under AI → External AI for owners and admins. From there they can review what each one can do, change its access, watch what it has been doing, and cut it off.
What a connection shows
| Field | Meaning |
|---|---|
| Assistant | The detected client, such as Claude, ChatGPT, Cursor, or VS Code |
| Connection name | The name you gave it when authorizing, so you can tell similar clients apart |
| Status | Connected or Disconnected |
| Widgets & channels | The inboxes this connection can reach |
| Permissions | The read and write capabilities it was granted |
| Reply as | The agent that replies and other write actions are attributed to |
| Connected | When the connection was authorized |
| Last used | When the assistant last called a tool |
Statuses
| Status | Meaning |
|---|---|
| Connected | The token is live and the assistant can call tools. |
| Disconnected | Access has been revoked. The record is kept so you keep the history. |
Manage access
Manage access opens everything chosen during sign-in. Only owners and admins can open and edit a connection.
Editable settings:
- Rename the connection
- Add or remove widgets and channels
- Add or remove read and write permissions
- Change which agent replies and other write actions are attributed to
Changes apply to the assistant's next request. The client does not need to reconnect or re-authorize, and removing a permission or a widget takes effect immediately.
Two rules still hold when editing: at least one widget or channel must stay selected, and at least one read permission must remain. Write permissions can only be granted by an owner or admin — see Permissions & access.
Latest activity
Each connection keeps its own audit log of the tools it has called, what they targeted, and when. Use it to answer "why did this conversation get resolved?" or to confirm an assistant is only doing what you expect before you widen its permissions.
Because replies are posted as a real agent, messages an assistant sends are also visible in the inbox like any other reply.
Usage and limits
Each connection is limited to 120 requests per minute, counted separately per connection, so one busy assistant cannot slow down another. The connection view shows how much of that budget is in use.
Sustained 429 responses on a single connection mean that assistant is polling too aggressively.
It is not a workspace-wide problem and does not affect your other connections or the REST API.
Multiple connections
You can connect more than one assistant at a time — for example Claude for triage and Cursor for development — and each client type gets its own token, widgets and channels, permissions, and usage budget on your team.
There is one connection per assistant client per team (one Claude, one Cursor, and so on). Use the connection name to tell similar setups apart when you use multiple custom integrations.
Only owners and admins see AI → External AI in Settings. They can connect, reconnect, change, or disconnect assistants there. Members do not have that menu and cannot complete MCP OAuth or edit connection settings.
Disconnect
Disconnecting asks you to confirm, then revokes the access token straight away. Only owners and
admins can disconnect from AI → External AI. The assistant
loses access mid-session and its next request fails with 401. The connection stays in the list
as Disconnected so you keep the audit trail.
Reconnecting the same client starts a fresh OAuth flow and issues a new token — the old one is never reinstated.
When access ends without you doing anything
- The team member whose token was issued is deactivated or removed from the team (their MCP token stops working; an owner or admin may reconnect for the team).
- Every widget and channel the connection pointed at has been deleted.
In both cases the assistant's requests start failing rather than silently falling back to wider access.
Real-time replies
The External AI page also shows whether webhooks are set up, since that is what lets your agent answer visitors automatically instead of only when you ask it to. If you have no webhook yet the page offers to create one; if you do, it links you to it.
Webhooks and MCP are independent — disconnecting an assistant does not stop your webhooks, and deleting a webhook does not affect MCP access. See the MCP overview for how the two fit together, and Webhook docs → Overview in the sidebar for setup.