MCP Server Overview

Prophia operates a hosted, remote MCP server that exposes your portfolio — accounts, properties, buildings, spaces, leases, documents, and the terms extracted from them — as tools an AI assistant can call. There is nothing to install or run on your side. You bring your own AI. Prophia provides the data connection, not the model or the chat interface. Any client that speaks the MCP Streamable HTTP transport works.

Two ways to connect

With OAuth, users sign in with credentials they already have, and there is no long-lived secret to distribute or rotate. If your goal is to connect Claude Desktop, you must use OAuth; if it is Copilot Studio, you must use an API key.

What your assistant can do

Ask questions in natural language and the assistant picks the tools. In practice that covers four things: Every answer traces back to a source document — tools return the page and the position on it, so the assistant can cite rather than paraphrase.

What it cannot do

The MCP server is read-only, with two exceptions: your assistant can manage custom terms and save portfolio reports. Everything else is read-only. If you need to write portfolio data programmatically, that’s the REST API, which has its own write access entitlement.

Permissions

Your Prophia permissions apply to every tool call. The same capabilities model that governs the REST API governs MCP:

• Connecting over OAuth gives the assistant at most what you can see. Building-level permissions filter the results, so two people at the same company can ask the same question and get different answers. On accounts that use role-based permissions, every OAuth connection is further limited to what you allow when you connect, whichever client you use.
• Connecting with an account API key gives the assistant admin-level access to the whole account, whoever is typing.
• Tools appear in the list whether or not you can use them. A tool you lack the capability for fails at call time with a message naming what’s missing — it isn’t hidden from the list.

Choosing what a connection can access

On accounts that use role-based permissions, every OAuth connection goes through a consent screen after you sign in. The client makes no difference: Claude Desktop, Claude Code, Cursor and every other MCP client get the same screen. It names the client and the Prophia account, and lists what the connection may read and, for custom terms and portfolio reports, change.

• Everything you can access starts selected. Uncheck anything you don’t want the client to use. A capability you don’t hold is greyed out: a connection can’t do more than you can.
• Allow Access requires at least one capability. Every tool call on the connection is then limited to what you allowed, on top of your own permissions. Account admins are limited the same way.
• Cancel stops the connection, and your client reports that access was denied.

Each connection gets its own choice, and Prophia doesn’t remember it for the next one. A client that registered with a client secret refreshes its access token without signing in again and keeps the choice until the connection reaches your account’s maximum connection lifetime. A client that registered without one signs in again when its 12-hour access token expires, and the consent screen appears again. To change what a client can access before then, disconnect and reconnect it. On accounts that don’t use role-based permissions there is no consent screen, and the connection gets everything you can see. API key connections never show it.

Built-in guides

The server publishes its own documentation as MCP resources — guide:// documents your assistant reads on demand to use the tools well. Assistants discover them with list_guides and fetch them with read_guides, usually without being asked. This is why the server tends to work well without prompt engineering on your side: the instructions travel with the tools.

Next steps