15. Connections from the inside

Chapter 7 described connections from the outside: belonging to the role or to the person, always through a single door, with the credentials out of the agent's reach. This chapter describes how that is built.

MCP: one door per role

The single door of chapter 7 is built with MCP (Model Context Protocol), the open standard for an agent to use external tools. For each role, the business definition exposes its own scoped MCP address: one that sees only what that role is allowed to use.

Behind that address is a gateway that presents the business's services as tools, whether the service speaks MCP or has an ordinary programming interface. The gateway reads the credential at the moment of each call, uses it and discards it; the model never sees it.

The business factory's conversations do not connect to that address directly: they do so through a local command that the coordinator mediates and signs, conversation by conversation. That way, each call goes out with the identity of the conversation that made it.

Credentials, in a vault

Each business's credentials live in a secrets vault, with one space per business and each environment (testing and production) kept separate. The engine that coordinates the agents never reads the credentials for external services or the logins to portals: the gateway uses them, or the provider that needs them stores them.

Browser connections

For systems that can only be used through a browser, the connection is a browser profile at a cloud browser provider:

  • there is one profile per connection to a site, holding only the login to that site (never a profile with all of someone's logins);
  • a role connection and a personal connection are separate profiles;
  • each conversation works with a disposable instance made from the profile, which does not save changes: the login cannot be altered from a conversation;
  • the instance is limited to the allowed sites, goes out through a connection in the appropriate country and closes on its own if unused.

One MCP gateway per role with the credential read at call time, and browsers with per-site profiles, disposable instances and a live view for the person.

The technical address used to control that browser is itself a credential, and that is why it is never handed to the model or to the business factory's machine. Two alternatives were also discarded: a profile with every site (a single stolen login would open everything) and action-by-action control from the coordinator (less capable than a full browser).

The live view

The login to each site is done by a person, by hand, in a live view of the browser. The same view serves to watch what an agent is doing and to take control: while the person has control, the agent cannot act in that browser.

The live view's address is also a short-lived credential: it is shown only to that person, on their screen, and never reaches the model or a log. Keeping the view available while an agent works costs considerably more than a browser without one, so it is off by default.

Outfitting, from the inside

A connection's loadout (chapter 7) takes shape in three places: the read and write tools are registered in the gateway, each with its level; the skills stay in the role's knowledge; and the routines are defined in the role's work and write their readings to its data store (chapter 14), so as to alert only when something changed.

To recap, MCP gives each role a scoped door, the gateway reads each credential at the moment of use and discards it, credentials live in a vault, and browsers work with per-site profiles, disposable instances and a live view for people only. The next chapter looks at how a business factory's files are organized.