OpenAI and Revit MCP: ChatGPT, Codex, and Local Connections

Connecting an OpenAI client to Revit MCP requires a local client that can launch Autodesk’s Windows server, a compatible Revit installation, and a verified model query. The vendor documentation describes the pieces of this route; this article has not hands-on tested their combination. OpenAI is the client provider in this setup, while Autodesk supplies the Revit connection.

Is there an official OpenAI MCP server for Revit?

Autodesk, not OpenAI, publishes the Revit Model Context Protocol (MCP) server as a local STDIO Technical Preview currently documented for Revit 2027.2, with tools grouped as Read, Write, and Experimental. The tool list can change, Read does not modify models, and support is version‑specific. Autodesk’s page names examples of local MCP clients. See Autodesk’s overview in Revit MCP server (Public MCP).

As of November 4, 2026, no OpenAI-owned Revit MCP server was identified in the reviewed documentation. Third-party projects may also use the phrase “Revit MCP”; check the publisher and requirements of the package you actually install.

Autodesk Product Help vs Revit model MCP

Autodesk also runs a separate remote MCP for Product Help. It is public, unauthenticated, and read‑only, intended for documentation search—not for querying or modifying your in‑session model. It is available as a ChatGPT integration and remains subject to workspace plugin policies. See Autodesk Product Help MCP.

Connect an OpenAI client to a local Revit MCP server

Use the following as a team procedure (a record you maintain internally), not UI instructions:

• Establish the Autodesk side: Record the exact Revit build (e.g., 2027.2 per Autodesk’s MCP docs), the MCP package used, and the server executable plus arguments as named in Autodesk’s documentation. Note the Read/Write/Experimental scope you expect to exercise.
• Choose a compatible OpenAI client: use the documented local MCP configuration in a supported desktop or Codex client, select STDIO, and supply the installed Autodesk executable and arguments. Confirm that the client can launch that Windows executable on the Revit host. Record its build and host; a cloud session or a Linux environment is not automatically equivalent.
• Keep one home for the config: Store a short, version‑labeled snippet in your internal IT notes or repo so teammates reproduce the same local settings without guesswork.
• Treat discovery separately from verification: First confirm the server starts and the client lists tools; then verify model access with a tiny, known selection before any writes.

For a plain‑language explainer of MCP boundaries and team hygiene, see Revit Model Context Protocol.

Why ChatGPT web and a Windows Revit server need different setup

ChatGPT on the web uses remotely hosted, MCP‑backed tools. A local Windows executable path is not a remote URL, so pointing the web app at a local STDIO server will not work. By contrast, MCP‑capable desktop clients (such as those documented by OpenAI) can connect to local STDIO or Streamable HTTP servers and share configuration per Codex host via app settings or a config file. Protocol support alone does not validate a particular Revit setup. See OpenAI MCP extension docs.

Same‑Windows restriction: Autodesk’s guidance requires the MCP‑capable client and Revit to run within the same Windows environment. Bridging from macOS to a Windows VM using gateways or adapters is not supported for the Revit MCP and transport reachability does not prove access to the live Revit session. See Autodesk support’s note on this constraint in Same‑Windows requirement for Revit MCP.

About tunneling: OpenAI’s Responses API can connect to remote MCP servers and supports a Secure MCP Tunnel for private servers, but this does not establish Autodesk support for tunneling the Public Revit MCP. Keep the Revit MCP local and within Autodesk’s documented boundaries. See OpenAI Responses API: tools, connectors, and MCP.

Verify model access with a small Revit query

Start with a precise, read‑only check that proves you are looking at the right document and the right elements. Matching a single count is not sufficient; linked models, element types versus instances, and phase or view scope can all mislead.

For a hypothetical first check, choose three native door instances in one test document. Record their element IDs and the Mark value shown in Revit, if that instance field exists. Inspect the server's available tools before deciding which fields to request; an unsupported field should remain unsupported.

Ask the client to query those specific IDs and return the document context, category, and chosen instance value for each. If the tool returns internal units for a numeric parameter in a later test, compare using an explicit unit convention rather than assuming its number matches the formatted Properties display. Keep the first test to a simple text field when possible.

Compare the exact returned ID set with the recorded set: no extras and no omissions. Check each value in Revit. Element IDs belong to a document, so a matching number without the correct document context is insufficient. File names can also be repeated across project copies; retain enough context to identify the actual test file.

If the result differs, reduce the query to one known element and inspect its tool response before changing the configuration. Distinguish a connection failure from a field-selection or scope error. Do not accept a fluent explanation in place of the requested native values.

A passing test establishes that this read worked in this environment. Record the prompt, tool response, versions, and native comparison so another team member can repeat it. Writes require their own supported task and controlled test. For broader product context, see Revit 2027 AI Review.

If your team needs Windows‑native assistance that can create/update elements in supported apps, consider ArchiLabs Desktop. Availability depends on application version, plan, accounts, and permissions.