AI assistants and the MCP server

Modero ships a Model Context Protocol (MCP) server. It lets an AI assistant, such as Claude, ChatGPT, or Gemini, read and change platform data on behalf of the person who connected it, in conversation, without anyone opening the web app.

A user connects once with their normal login. From then on they can ask things like:

Which contracts expire in the next 60 days?

Show me every service at the Dallas office.

Open a ticket for the circuit that keeps flapping and assign it to me.

The one thing to tell anyone who asks#

The assistant can never see or do more than the person who connected it. It inherits their exact permissions: their organization, their partner, their role. There is no separate AI account with elevated access, and there is no way to grant one.

Nearly every question about MCP resolves against that rule. If a user could not open a record in the web app, their assistant cannot reach it either. If their role changes, what MCP can reach changes immediately, because access is evaluated per request from their live role and memberships.

One server, both brands#

Modero and its branded instances are the same application behind different brands, so there is one MCP server for all of them. Vigilis, the reference branded instance, connects to the same server your advisory firm does.

What differs between two people is their reach, and that comes from their role, not from which brand they signed in through:

An OAuth token carries no notion of a “currently selected organization”, so partner staff see all of their reachable client organizations at once. Naming one narrows it: most tools accept an organization, and in practice you just say “for Acme Manufacturing” and the assistant resolves it.

What it is not#

• Not a data export. Every result is bounded and filtered to the caller’s access.
• Not an admin backdoor. User, organization, and partner administration is not exposed over MCP. Those stay in the Partner Hub.
• Not automatic. Nothing happens unless someone asks for it in a conversation.
• Not a replacement for the product. It reaches the same records, through a different door.

Where to go next#

• Turning MCP on covers the four things that must all be true before a connection returns any data.
• Connecting an assistant is the walkthrough you can send to a user.
• What an assistant can do is the tool catalog, for answering “can it do X?” precisely.
• Changing data safely covers previews, protected fields, and bandwidth change orders.
• Governing connected assistants covers revoking access and capping a client’s assistant at read-only.
• Getting good answers is the habits and prompt patterns worth passing on to a client.
• When MCP is not working is ordered by how often each cause turns out to be the real one.

Common questions#

Does the AI get my password? No. Authentication is OAuth 2.1 with PKCE against Modero itself. The AI vendor receives a scoped, expiring token, never a credential.

Can an assistant see another customer’s data? No. Every query is filtered by the same tenancy rules the web app uses, evaluated per request.

Can it see stored secrets or API keys? No. Columns whose names indicate credentials are stripped from every read, and a request that names one explicitly is refused rather than silently dropped.