Skip to content

Revit Model Context Protocol

Author

Brian Bakerman

Date Published

AI Agents for Architecture

Revit Model Context Protocol, usually called Revit MCP, connects an AI application to tools that can read or change a Revit model. The assistant interprets your request; the MCP server provides the supported operations; Revit holds the actual project data.

That can make a request such as “find the doors with missing asset references” useful inside a live project. It does not automatically give the assistant every Revit command or access to every linked document.

Updated November 10, 2026. Autodesk now documents separate Read, Write, and Experimental public MCP server options. This guide also covers Claude and OpenAI client choices, community projects, and ArchiLabs Desktop.

What is Revit Model Context Protocol?

MCP is an open standard for connecting AI applications to external tools and data. “Model” in its name refers to the AI model; MCP itself is not a BIM file format, a replacement for IFC, or a system that automatically synchronizes project files. Official MCP introduction.

In a Revit workflow, the pieces have different jobs:

  • The AI application contains the conversation and decides which available tool to request.
  • The Revit MCP server exposes specific operations, with their inputs and outputs.
  • Revit and its add-on carry out the supported work in the appropriate document.

For example, a query tool might return door IDs, a data tool might retrieve their properties, and a write tool might update an editable parameter. If the server provides only the first two, asking for the third does not make it available.

Autodesk's Revit MCP server: Read, Write, and Experimental

Autodesk's current public MCP overview describes a Technical Preview with a local STDIO connection and three server configurations. Read provides queries without model modification; Write adds supported changes; Experimental includes further developing tools. The overview and setup pages specify Revit 2027.2, so confirm compatibility against the installer and documentation for the build you use. Autodesk Revit Public MCP Server.

The tool reference makes the differences concrete. It lists parameter changes, moving, and rotating elements under Write; several sheet, view, and schedule operations remain under Experimental. Read capabilities include model queries, element data, navigation, and exports. Use the current tool list when planning a task, because the preview can change. Autodesk's tool reference.

This is more informative than labeling every Revit MCP server “read-only” or assuming every one can create a drawing package. Ask which implementation and configuration the person is using.

Autodesk Assistant inside Revit is another product surface. Its supported operations are documented separately from the public MCP add-on. A demonstration in Assistant does not establish that the same command is available in your chosen public server configuration.

Connecting Claude Desktop to Revit MCP

Autodesk provides a Claude Desktop setup example. It requires the relevant Revit MCP add-on, a supported Revit installation with a model open, and Claude's local server configuration. After configuration, the client must be restarted. Use Autodesk's current example for the installed server rather than copying an executable path from an older tutorial. Official Claude Desktop setup instructions.

Begin with two checks: list the available tools, then identify the active Revit document and view. These establish what is connected before you ask questions about project content.

A useful first task is to list a small, known set of elements and compare their returned values with Revit's Properties palette. For example, inspect the Mark values of three selected doors. This is easier to diagnose than a project-wide query that produces hundreds of records.

OpenAI, ChatGPT, and Revit MCP connections

An OpenAI model and an MCP connection are separate parts of a workflow. OpenAI's API supports remote MCP servers and private or local servers through Secure MCP Tunnel. A developer still needs to supply a Revit server and the surrounding application behavior. OpenAI's MCP documentation.

For ChatGPT on the web, OpenAI documents custom MCP plugins using a server URL or tunnel, subject to workspace permissions. Desktop-only plugins can use local servers through a separate supported route. Check the actual ChatGPT surface and installed plugin; an example written for one client is not a universal setup file. ChatGPT custom MCP guide, desktop-only plugins.

These connection options should not be described as an OpenAI-authored Revit integration unless the specific integration is identified and verified. Our ChatGPT for Revit guide explains the difference between script assistance, document review, and a live model workflow.

Community Revit MCP tools and pyRevit

Community servers provide another route into Revit, sometimes through a custom add-in and sometimes through pyRevit. Their installation requirements and exposed operations vary, so read the project's own documentation rather than treating “Revit MCP” as one download.

The pyRevit project also documents an agent runtime under active development. Its status is experimental, it is off by default, and the documented host validation covers Revit 2024 and 2025 while other phases remain in progress. This is a development effort to watch, not a blanket statement that every stable pyRevit installation includes a finished MCP assistant. pyRevit agent runtime documentation.

If your team already maintains pyRevit tools, evaluate how a server fits that environment. Who maintains it after a Revit update? Which commands can it run? How will a designer inspect a change? These practical questions matter as much as the protocol.

What ArchiLabs Desktop adds around a Revit connection

ArchiLabs Desktop provides an AI workspace for native Revit and AutoCAD work. It combines model choices from OpenAI and Anthropic with project files, annotated screenshots, selected native elements, narrated recordings, and editable knowledge libraries.

An MCP server solves the connection part of a task. A working project assignment also needs the brief, source material, instructions, outputs, and follow-up conversation. Desktop brings those parts together for the designer.

For example, attach an Excel equipment register and a marked-up PDF, capture the relevant Revit selection, and explain the intended update. Keep the accepted naming and source-priority rules in project knowledge so a colleague can reuse them. The public equipment demonstration shows a PDF becoming a native Revit family and a placed instance.

Choose a standalone MCP setup when you want to assemble and maintain your own AI workflow. Evaluate Desktop when you want the surrounding CAD and BIM working environment provided as a product.

How to test a Revit MCP workflow

Use a small model or a copy of a project and test in stages:

  1. Confirm the connection. Identify the client, server configuration, active document, and available tools.
  2. Check a query. Compare returned element IDs and values with known objects in Revit.
  3. Check the scope. Establish whether the task includes the whole document, a view, a selection, or linked data.
  4. Try one supported edit. Specify the exact target and value, then inspect the native result.
  5. Review the handoff. Keep the output and any unresolved items so another person can understand what happened.

A locally running server does not by itself keep all model information local: data returned to an AI client may be sent to its model provider. Review the actual client and server configuration against your project's information requirements.

For a broader guide to native AI work, read Revit AI: features, tools, and practical workflows. Start with a task your team can inspect, then expand as the connection and results become familiar.