MCP Overview

Connect agents to external systems with Model Context Protocol.

MCP lets Comms agents read external systems through structured tools.

Use MCP when an agent needs live or private data:

• Bookings.
• Orders.
• Customer profiles.
• Memberships.
• Support tickets.
• Finance schedules.
• Repository metadata.

Add an MCP server

In the dashboard:

1. Go to Sources.
2. Add an MCP server URL.
3. Configure authentication headers or API keys.
4. Discover tools.
5. Choose an access policy.
6. Add the MCP server to an agent.

Access modes

Comms supports three MCP access modes:

Mode	Use case
admin_only	Owner/admin runs only. Never exposed to subscriber conversations.
subscriber_safe	Subscriber conversations can use only safe, read-only, scoped tools.
public	Public read-only tools may be exposed when they do not reveal private broad records.

For customer-specific private data, use subscriber_safe.

Tool discovery

Comms stores discovered tool schemas on the MCP server record. During an agent run, the runtime filters those schemas based on:

• Whether the server is enabled.
• The current actor type.
• The access policy.
• Tool annotations and names.
• Whether the tool looks read-only and subscriber-scoped.

Tools that do not pass policy are not shown to the model.