Revit MCP - AI-Powered Revit Control

English | 繁體中文

Revit MCP lets AI clients call Autodesk Revit tools through the Model Context Protocol (MCP). The MCP server forwards tool calls to a local Revit add-in, and the add-in executes the corresponding Revit API workflow.

• Demo video: Revit MCP - AI-Powered BIM Workflow Demonstration
• Knowledge site: https://shuotao.github.io/REVIT_MCP_study/
• Default WebSocket port: 8964

What is this?

Talk to Revit in plain language. Ask your AI client to "dimension every wall on this view" or "check the curtain-wall elevations", and Revit does it — through 192 MCP tools backed by 83 professional BIM SOPs (building code, quantity take-off, compliance checks).

Who it's for: BIM engineers and architects who use Revit and want AI-assisted, standards-based workflows. You'll need Revit (2022–2026) on Windows and to be comfortable installing an add-in.

Quickstart (3 steps)

1. Install the Revit add-in. Build and deploy the C# add-in — see Manual Setup. This is the half that actually talks to Revit.
2. Point your AI client at the MCP server. No cloning needed — it runs straight from npm:
   config.json

   Copy

   { "mcpServers": { "revit-mcp": { "command": "npx", "args": ["-y", "@shuotao/revit-mcp-server"] } } }
   

3. Open Revit, enable the MCP service in the ribbon, and start asking. Full setup: AI Client Configuration.

Questions or want to show what you built? → Discussions

Current Project Counts

When these numbers change, update CLAUDE.md, README.zh-TW.md, this file, docs/DOCUMENT_AUDIENCE_INVENTORY.md, and run:

.\scripts\verify-qaqc.ps1 -SkipBuild -SkipDeploy

New: Interactive Clash Viewer (MCP Apps)

detect_clashes can now render an interactive clash-review UI inline in the conversation, via the MCP Apps extension (io.modelcontextprotocol/ui). This requires an MCP host that supports the extension. It is purely additive:

• Interactive rendering only happens in GUI hosts that support MCP Apps (Claude Desktop, the claude.ai web app, VS Code GitHub Copilot). A terminal CLI like Claude Code can't render the panel and always shows the text result — this is expected, not a fault.
• Hosts without MCP Apps support keep getting detect_clashes's normal text result — nothing changes for them.
• The stdio connection and the existing Revit add-in are unaffected; no reinstall or reconfiguration is needed.

See docs/MIGRATION_GUIDE.md for the full MCP 2026-07-28 upgrade notes.

Architecture

AI Client
  Claude Desktop / Claude Code / Gemini CLI / VS Code Copilot / Antigravity
        |
        | stdio
        v
MCP Server
  Node.js / TypeScript
  MCP-Server/build/index.js
        |
        | WebSocket ws://localhost:8964
        v
Revit Add-in
  C# / Revit API
  MCP/Application.cs
  MCP/Core/SocketService.cs
  MCP/Core/ExternalEventManager.cs
        |
        v
Autodesk Revit

External AI clients do not need an API key inside this repository. Their account and authorization are managed by the AI client itself. Only an embedded Revit chat feature that directly calls an AI API would need an API key.

Requirements

One-Click Setup

Recommended for new users:

.\scripts\setup.ps1

For AI agents or non-interactive setup:

powershell -ExecutionPolicy Bypass -File scripts/setup.ps1 -NonInteractive -RevitVersions "2024,2025"

The setup script checks prerequisites, installs dependencies, builds the MCP server, builds and deploys the Revit add-in, and helps configure common AI clients.

Install from MCP Registry

The MCP server is published on the MCP Registry as io.github.shuotao/revit-mcp-server (npm package @shuotao/revit-mcp-server). Run it directly with:

npx -y @shuotao/revit-mcp-server

Note: this npm package is only the Node stdio bridge. The C# Revit add-in (MCP/) must still be installed separately — follow Manual Setup below for your Revit version.

Manual Setup

1. Build the MCP Server

cd MCP-Server
npm install
npm run build

AI clients launch:

node MCP-Server/build/index.js

2. Build the Revit Add-in

Choose the configuration that matches your Revit version:

cd MCP
dotnet build -c Release.R22 RevitMCP.csproj   # Revit 2022
dotnet build -c Release.R23 RevitMCP.csproj   # Revit 2023
dotnet build -c Release.R24 RevitMCP.csproj   # Revit 2024
dotnet build -c Release.R25 RevitMCP.csproj   # Revit 2025
dotnet build -c Release.R26 RevitMCP.csproj   # Revit 2026

Expected output:

MCP/bin/Release.R{YY}/RevitMCP.dll

Example for Revit 2024:

MCP/bin/Release.R24/RevitMCP.dll

3. Deploy the Add-in

Recommended:

# Pick the Revit version explicitly. With several installed and no -Version, the script asks;
# in -NonInteractive mode it fails rather than guessing (it never silently picks the highest).
.\scripts\install-addon.ps1 -Version 2024

# Deploy to every installed version that has a matching build
.\scripts\install-addon.ps1 -All

The script copies every DLL in the build output (13 for R22-R24, 8 for R25-R26 — both are correct), checks the build generation matches the target Revit before copying, verifies each file by SHA256 afterwards, and rotates old backups.

For manual deployment, place the .addin file and DLL under the matching Revit Addins directory, and keep the relative assembly path in RevitMCP.addin:

<Assembly>RevitMCP\RevitMCP.dll</Assembly>

Do not create version-specific .addin files, and do not hardcode absolute DLL paths.

AI Client Configuration

Project-level .mcp.json:

{
  "mcpServers": {
    "revit-mcp": {
      "type": "stdio",
      "command": "node",
      "args": ["./MCP-Server/build/index.js"],
      "env": {}
    }
  }
}

VS Code config in .vscode/mcp.json:

{
  "servers": {
    "revit-mcp": {
      "type": "stdio",
      "command": "node",
      "args": ["${workspaceFolder}/MCP-Server/build/index.js"],
      "env": {}
    }
  }
}

Other AI clients use the same concept: launch MCP-Server/build/index.js with node.

Config template per client:

Replace <YOUR_PROJECT_PATH> in the templates with the actual project path on your machine.

Switching Between AI Clients

The Revit-side WebSocket service holds an exclusive lock: only one AI client can be connected at a time. A second client that tries to connect is cleanly rejected (HTTP 409), not swapped in — the first connection stays stable. Multiple AI clients are therefore used by switching, not concurrently:

Before that lock check even runs, any handshake carrying a browser Origin header is rejected outright with HTTP 403. This closes a separate risk: a malicious page open in your browser could otherwise drive the add-in without your knowledge, since WebSocket handshakes aren't covered by the browser's same-origin policy. It has no settings opt-out and does not affect the MCP bridge, which never sends an Origin header.

1. In the Revit ribbon, click the "切換/釋放連線" (Switch/Release Connection) button to release the current connection.
2. Start or reconnect the other AI client; once its MCP server connects to localhost:8964, it takes the lock.
3. The "MCP 設定" dialog shows which client currently holds the connection (e.g. claude-code, claude-ai).
4. If the connection misbehaves, use the same ribbon button, or restart the MCP service from the Revit ribbon to reset it.

Startup Flow

1. Start Revit.
2. Open or create a Revit project.
3. Enable the MCP service from the Revit ribbon.
4. Confirm the Revit add-in is listening on localhost:8964.
5. Start or restart the AI client so it loads the MCP server.
6. Call Revit MCP tools from the AI client.

If localhost:8964 is unreachable, Revit may not be running, the MCP service may be off, the port may be occupied, or the AI client and Revit add-in may be using different port settings.

Project Structure