What is MCP? A plain-English guide for BIM professionals

If you have watched someone ask Claude to count the doors in a Revit model, you have seen MCP at work. The term comes up in every AI and BIM conversation now, usually without an explanation. Here is the short one.

The one-sentence version

MCP, the Model Context Protocol, is an open standard that lets an AI app use tools on your computer, such as reading your Revit model or creating a wall in Archicad, through a small program called an MCP server.

Anthropic published it in November 2024. Since then it has been adopted well beyond Claude, including by OpenAI for ChatGPT. That is why the same connector can work with more than one AI app.

The three pieces

Think of it as a chain with three links:

1. The AI app. Claude Desktop, Claude Code, ChatGPT. This is where you type.
2. The MCP server. A small program running on your PC. It tells the AI app which tools exist, for example “read elements”, “run a script”, “capture a view”, and carries out the calls.
3. Your software. Revit, Archicad or anything else with an API. The server talks to it directly on your machine.

You  ->  AI app  ->  MCP server  ->  Revit / Archicad

When you ask “Which ducts have no insulation?”, the AI decides which tool to call, the server asks Revit, and the answer comes back to the chat.

What runs where

This part matters for your projects:

• The MCP server runs on your PC. ArchiBridge, for example, only talks to Archicad on 127.0.0.1, the address of your own computer.
• The AI model runs in the cloud. Whatever the AI reads to answer you (element names, parameters, view captures) is sent to the AI provider, Anthropic or OpenAI.

So “local” does not mean “private”. If a project has strict confidentiality rules, check them before you connect a live model, and try everything on a copy first.

Reading versus changing

Good connectors keep reading and changing apart. ArchiBridge has separate tools for reads and modifications, so your AI app can ask you before any change. BIMBridge runs changes as scripts inside Revit and asks you to Keep or Discard each one, and a kept change undoes with one Ctrl+Z.

Use that separation. Start read-only, check the answers against what you see, and only then allow changes.

What MCP cannot do

• It does not make the AI right. The AI can still misread a request. Specific prompts and a quick check catch most of it.
• It only reaches what the software exposes. If Revit’s API cannot do something, neither can the AI.
• Big models are slower. Asking about 200,000 elements takes longer than asking about one level. Narrow the question to a level, a view or a system.

Try it

Two free connectors get you started:

• BIMBridge connects Claude and ChatGPT to Revit through pyRevit’s MCP server.
• ArchiBridge does the same for Archicad, through the Tapir add-on.

Both install with one prompt. The guide How to connect Claude to Revit walks through the first session.