MCP Server and Agent Tools

The MCP concept opens the portfolio to AI clients that live outside Subscriber Bot. Instead of a general-purpose integration that can do anything, the design publishes a catalog of named tools — list relationships, fetch one by identifier, create a tracked relationship, list inbox messages, change an agent's posture, attach a policy, activate a draft workflow — each declaring what it does, what it needs and whether it only reads.

Connecting an assistant to financial and contractual records is exactly where a vague integration becomes dangerous. A tool that cancels a subscription and a tool that lists subscriptions should not look alike in a client's tool list, and whether a call requires confirmation should be visible before it is made rather than discovered afterwards. This console makes those distinctions part of the catalog itself.

The illustrated view is built around that idea. A header shows the server, its connection state, the transport in use and the endpoint it is pointed at, with the identity the session is acting as on the right. Three tiles count the domain tools, the platform modules behind them and — highlighted as an attention state — the tools classified as destructive. A search field and read-only, write and destructive filters narrow a long catalog quickly.

The catalog itself is a table with a column for the tool name, the area it belongs to, whether it reads or writes, whether confirmation is required, and a plain description of what it does. Destructive rows are tinted and marked with a warning icon, so cancelling a relationship, archiving a provider and removing a payment instrument stand apart from the reads around them. Selecting a row opens a detail panel with the tool's purpose, its required inputs and their types, and the result the last call returned — including error text, so a failed call is diagnosable from the same panel.

A warning strip along the foot of the page flags that confirmation is not being required for the destructive tools, which is the point of the whole design: the governance state is shown rather than assumed. The catalog mirrors the same posture and policy thinking that governs the built-in agents, extended to whatever client is connected.