How to connect AI to Autodesk Maya: Claude, ChatGPT and Cursor explained simply

This guide is for animators, modelers, riggers and 3D generalists. We explain what to install, where to enter commands and how to tell whether the assistant is actually working with your scene. Instructions were checked against project documentation on 4 November 2026. This is a documented setup route, not a report of tests on every Maya version: compatibility must be checked on your own computer.

What does “connect AI to Maya” actually mean?

There are three different workflows.

MCP is a standard that lets AI call tools in another application. A Maya MCP server is a small bridge between the chat and Maya. It neither replaces Maya nor acts as a separate AI model.

The practical distinction is simple: a script written in chat has not changed your scene yet. A tool call, a response from Maya and checked object values are evidence of actual work.

The simplest start: run a script directly in Maya

Yes: your first useful operation can work without MCP, a terminal or plugin installation. ChatGPT, Claude or Cursor chat writes Python, and you run it in Script Editor. This is a manual exchange: the chat itself cannot see the open scene or press Run.

1. Save the scene and work on a copy.
2. Tell the assistant your Maya version, select the relevant object and describe one specific operation.
3. Open Windows → General Editors → Script Editor and choose the Python tab.
4. Paste only the code, without triple backticks or explanations. Review it before running it.
5. Select the entire required block and execute it. Copy the output or complete error message back into the chat.

A ready-to-use prompt for a first attempt:

Write a Python script for my Maya version using only maya.cmds. Read the selected objects, current frame and scene units, then print a clear report. Do not modify, save or download anything. Explain the code before I run it. Do not claim you have already executed it.

Here is a short example report you can paste directly into the Python tab:

import maya.cmds as cmds

print("Maya:", cmds.about(version=True))
print("Frame:", cmds.currentTime(query=True))
print("Linear unit:", cmds.currentUnit(query=True, linear=True))
print("Selected:", cmds.ls(selection=True, long=True) or [])

This example only reads data. Give its output to the assistant and request the next script using actual object names. Add a screenshot for pose assessment; motion needs frames or a playblast in a format your chat supports. Without the returned result, the assistant cannot know what happened in Maya.

One Shelf button instead of pasting repeatedly

Once a script has been checked, save it as a button: select the required Python code in Script Editor, choose File → Save Script to Shelf and give it a clear name. This is a standard Maya feature described in the Autodesk instructions.

You can save either your own tool or the local-port opening code from step 2 below. After the MCP client has been configured once, daily startup becomes simpler: open Maya, press the button and check the connection in chat. The port button alone does not install MCP or connect your account to a model.

Easier Claude setup: a packaged MCPB extension

The community project GG_MayaMCP documents a packaged Claude Desktop extension that avoids manual JSON editing. This is a separate alternative to the main walkthrough below.

1. Open project releases and look for maya-mcp-version.mcpb under Assets. Without a packaged file, this shortcut is unavailable for that release.
2. In Claude Desktop, use Settings → Extensions → Advanced settings → Install Extension. Local extensions must be allowed.
3. Keep port 7001 and leave arbitrary script.run execution disabled.
4. Review scripts/enable_commandport.py from the same version and run the complete file in Maya's Python tab. It handles Maya 2024 compatibility separately; do not substitute another bridge's script.
5. Request health_check, scene_info and nodes_list. Compare the response with your scene.

The modeling_create_polygon_primitive tool can create a test cube without enabling arbitrary Python. This package targets desktop Claude with Maya on the same computer, not ordinary browser chat.

Documentation availability does not mean our team tested the package in a live Maya session. We could not retrieve the releases listing during preparation, so we cannot confirm a particular current download. Check its source and version requirements, and use a test scene.

Recent videos: AI working inside a Maya scene

These examples concern connecting an assistant to Maya, not just obtaining code in chat. They are third-party demonstrations; test commands on a copy of your scene and follow your studio's rules on data shared with a model.

1. Claude Code works with Maya without MCP

Chris Cook, “Claude in Maya, no MCP!”, 16 October 2026, 11 min 39 sec. The creator describes Claude Code working in Maya, from renaming objects to rig fixes. A section at 9:31 covers fixing a shoulder seam. The method requires Maya 2022 or later and Claude Code running locally; it uses the creator's own bridge rather than MCP.

Open Chris Cook's video on YouTube

2. Setting up Maya MCP and Claude Opus

fx node, “Maya MCP and Claude Opus part 1”, published approximately October 2026, 9 min 13 sec. The creator documents an attempt to set up MCP in Maya. This illustrates the connection process and does not establish the quality of automatic animation review.

Open fx node's video on YouTube

Titles, creators and descriptions were checked on YouTube. The browser could not play the videos during verification, so this account relies on the creators' descriptions and chapter lists rather than a complete viewing. If an embedded player is unavailable, use the link below it.

Which approach should you choose

If you want to talk to the assistant in a familiar chat window, start with Claude Desktop and a local MCP server. If you also build scripts, Shelf buttons or team tools, Cursor in Agent mode.

Cursor is an AI-powered editor, not a separate model. You choose an available model inside it, while the Maya connection stays the same. Having Claude in Cursor does not make Cursor and Claude subscriptions interchangeable.

The walkthrough below uses the open-source community project mcp-for-maya. It documents tools for reading scenes, running Python, performing checks and capturing the viewport. This is a community project, not an official Autodesk plugin. It was selected for capabilities useful in this tutorial, not on the basis of comparative reliability testing.

Before starting: a separate test scene

Save your working file. For this tutorial, open an empty scene and save it as ai_maya_test_v001.ma. Do not start with a client rig, complex references or the only copy of a shot.

The bridge can execute Python with your Maya user permissions. Autodesk explicitly warns that commandPort has no built-in authentication. Use a local connection only: do not forward the port on your router or expose it to the internet. Local execution does not make arbitrary Python safe either: install only reviewed packages. Autodesk documentation.

Data returned by the server may reach the model provider: object names, paths, parameters and scene captures. Check studio rules before using studio assets.

Step 1. Install uv outside Maya

uv helps run Python tools in a separate environment. Install it using the official Astral instructions. On Windows, you can open PowerShell and run:

winget install --id=astral-sh.uv -e

After installation, close and reopen your terminal. Check:

uv --version
uvx --version

Enter these commands in your operating system’s terminal, not Maya’s Script Editor. This approach does not require installing the MCP package into Maya’s embedded Python.

Step 2. Enable a local connection in Maya

Open Windows → General Editors → Script Editor. Select the Python tab, paste this code and execute it:

import maya.cmds as cmds

port = "127.0.0.1:7001"
if not cmds.commandPort(port, query=True):
    cmds.commandPort(name=port, sourceType="python")
print("Maya connection:", cmds.commandPort(port, query=True))

The expected result is True. The address 127.0.0.1 means this computer. The number 7001 is the connection port. If another process occupies it, do not close that process blindly: choose another unused port and use the same number when connecting the client.

Do not add automatic startup to userSetup.py until the first test succeeds. First, establish a simple connection that you know how to turn on and off.

Step 3A. Connect Claude Desktop

You need the desktop application. Find Developer → Edit Config in its settings. If your version looks different, consult the local MCP guide.

Add the server to claude_desktop_config.json:

{
  "mcpServers": {
    "maya": {
      "command": "uvx",
      "args": ["mcp-for-maya"],
      "env": {
        "MAYA_MCP_INCLUDE_PORTS": "7001"
      }
    }
  }
}

If the file already contains other servers, add only the maya entry inside the existing mcpServers object; do not erase the rest. Fully quit Claude Desktop and reopen it. The first launch may download the package and its dependencies.

Check whether the server and its tools appear. A configuration entry alone does not prove that Maya is connected.

Step 3B. Or connect Cursor

The uv and Maya steps stay the same. Open an experiment folder in Cursor, create `.cursor/mcp.json` inside it and paste the JSON above. For all projects, you can use the global ~/.cursor/mcp.json.

Restart Cursor, check the server status in MCP settings and open an Agent chat. Keep action confirmations enabled while learning. Configuration and file locations are covered in the Cursor documentation.

If the client cannot find uvx, run where.exe uvx on Windows or which uvx on macOS/Linux. You can put the resulting absolute path in the command field. In JSON, write Windows paths with doubled backslashes or ordinary /.

Step 4. Test the connection without modifying the scene

Send the assistant:

Connect to Maya using the available MCP tools. First call list_sessions. If automatic discovery finds nothing, try add_session with host 127.0.0.1 and port 7001. Report the Maya version, open scene name, current frame and selected objects. Change nothing. If the connection fails, show the actual error.

Compare the answer with your open Maya session. For a stronger test, manually create a cube with an unusual name, select it and repeat the request. The assistant should read that exact name rather than give a generic “I’m connected” response.

Then give it a small first task:

In the test scene, create a cube named AI_TEST_CUBE with sides of 10 current linear units. Do not change the scene units. If that name already exists, stop. Read back the actual dimensions and report the result. Do not save the file automatically.

Check the Outliner and geometry. This is a basic test of the “read → execute → verify” cycle.

How to make AI check its own work

Request two kinds of checks. Technical checks read actual data: names, keys, transforms, polygon counts and connections. Visual checks use a viewport capture, render or playblast frames, provided the client can receive and inspect images.

A text log cannot show whether a character’s pose works. One screenshot cannot show whether a foot slides between frames. And the right number of keys does not guarantee convincing animation.

Add this rule to your tasks:

Before working, read the scene state and explain a short plan. After each change, read the changed data again. For a visual conclusion, obtain and inspect an image of the result. If you cannot inspect images or motion, say so directly. Do not describe anything as verified unless you checked it.

Tasks for an animator

Start with a specific controller and a short range:

Analyze keys only on the selected controller between frames 1–48. Report animated channels and value ranges. Find suspiciously abrupt changes between adjacent frames and explain each flag. Do not modify keys, animation layers, constraints or references.

After reviewing the report, you can approve a separate correction. For example, change one agreed key in a scene copy and compare motion before and after. For foot contact, request world-space measurements and inspection of multiple frames. Avoid starting with “make all the animation better”: the criteria are too vague.

Tasks for a modeler

Create a simple table blockout in a separate AI_TABLE_TEST group: a tabletop and four legs. First read the scene units and propose dimensions in centimeters. After my confirmation, build the geometry and check its dimensions and names. Leave existing objects untouched. Show top and perspective views.

For an existing model, begin with an audit: request a list of issues and select the objects needing attention. Approve vertex merging, history deletion, remeshing or UV changes separately, particularly when a model is already rigged.

Tasks for a rigger or technical artist

Inspect only the selected controllers: locked channels, incoming connections and constraints. Produce a table. Do not disconnect, rename or modify reference edits. Suggest one safe test for a suspicious area.

Another practical direction is building your own tools. Ask Cursor for a small panel that performs a repetitive operation, explains its code and tests it on dummy objects first. The result is a reusable button you can use without starting another conversation.

What if you specifically want ChatGPT?

For ChatGPT in a browser , Maya’s local address is not a ready-made connection. You need an MCP bridge and a supported access path. The current OpenAI instructions describe adding custom servers through Plugins; availability depends on your account and workspace permissions.

One route that avoids publicly exposing Maya is Secure MCP Tunnel. The sequence is:

1. Verify the local Maya connection first.
2. Configure a tunnel and the required permissions in Platform.
3. Install the official tunnel-client.
4. Configure it to run the MCP server command uvx mcp-for-maya, run diagnostics and leave the client running.
5. Add a custom MCP server in ChatGPT, choose Tunnel and select the appropriate tunnel.
6. Check the tool list and perform a read-only scene test.

Do not enter 127.0.0.1:7001 as the MCP server URL: it is Maya’s command port, not an HTTP MCP endpoint. If your account lacks the required options, this route needs additional access setup. Claude Desktop or Cursor require less infrastructure for a first local experiment.

Chat inside Maya and other bridges

There are also panels that embed chat in Maya and call a model API. For example, ChatGPTforMaya is a real project, but its repository was archived on 15 May 2026. An old video tutorial is not a guarantee of compatibility with current Maya versions and APIs.

Other MCP projects include maya-mcp-server, MayaMCP, and also dcc-mcp-maya. Their installation commands, tools and launch methods differ. DCC-MCP, for example, has its own adapter ecosystem and CLI workflow. Do not combine configurations from different repositories into one setup.

If something does not work

To close the connection created in this guide, run this in Maya’s Python tab:

import maya.cmds as cmds

port = "127.0.0.1:7001"
if cmds.commandPort(port, query=True):
    cmds.commandPort(name=port, close=True)

What does it cost, and where should you start today?

An open-source MCP bridge does not remove Maya licensing, client subscription terms or model usage costs. If a plugin asks for an API key, check its billing separately and set a limit. Do not assume that an ordinary chat subscription automatically covers every API plugin.

A good first result is not “a finished film from one prompt,” but a small, verifiable operation: reading controller keys, building a blockout, finding naming issues or preparing a button for repetitive work. Once that cycle is reliable, gradually expand the tasks. Artistic decisions, deformation quality and final shot approval remain the specialist’s responsibility.

How to use Cursor →

AI for 3D animation and filmmaking →