
Claude and Codex for CAD: SolidWorks, AutoCAD, CATIA
TL;DR
Neither Claude nor Codex is a universal winner for CAD. What decides the result is the integration that reaches your application and how you verify the saved file. Start with one bounded task on a copy, in both assistants, and compare the verified outcome.
Updated
·
6 chapters
·
11 min read
Start here: read the first post →
See the chapters ↓
In this guide
What you’ll learn
6 chapters

01
Claude or Codex: pick the workflow, not the assistant
Why the model, the client and the CAD connection are three separate things, and how to compare fairly.
Jump to chapter ↓

02
SolidWorks with Claude and Codex
The two twin guides: inspect, change and verify a part, assembly or drawing through MCP and computer use.
Jump to chapter ↓

03
AutoCAD with Claude Code and Codex
Live drawings through an MCP connector, AutoLISP, and the limits of each connector.
Jump to chapter ↓

04
CATIA with Claude and Codex
A community CATIA V5 bridge, why V5 and 3DEXPERIENCE are not interchangeable, and what to test first.
Jump to chapter ↓

05
Revit, Fusion, Rhino, SketchUp, Inventor, FreeCAD and KiCad
The other engineering and design apps that have a Claude guide, a Codex guide, or both.
Jump to chapter ↓

06
Run the CAD app and the agent on one Windows machine
Most CAD connectors need Windows and the app on the same host. Here is how that works from a Mac or a weak laptop.
Jump to chapter ↓
If you do CAD and want an AI assistant to do some of the repetitive work, the first question is usually "Claude or Codex?" The better question is what can reach your CAD application, and how will you check what it did. This guide puts the Claude and Codex posts side by side for SolidWorks, AutoCAD and CATIA, then covers the other CAD and design apps. It ends with a fair test you can run on your own files. It summarizes what Vagon's guides document and is not a benchmark.
Claude or Codex: pick the workflow, not the assistant
Vagon's Claude vs. Codex for CAD guide starts from a plain conclusion: neither assistant is a universal winner across SolidWorks, AutoCAD and CATIA. The outcome depends on three layers that people tend to blur together.
The model turns your request into code or tool calls. The client (a terminal session, a desktop app) decides which files, commands and integrations are available. The connection to the CAD application decides what actually reaches the model. If Claude gets a working model-query tool and Codex only gets screenshots, you are comparing two setups, not two models.
There are three ways an assistant can touch CAD. It can write a script that you review and run. It can use an MCP server that exposes operations. Or it can use computer use to click through the interface. MCP vs. Computer Use for Creative Apps gives a practical order: use an existing script or command-line tool if it already does the job, then an MCP tool, then computer use, and combine them only where one leaves a gap.
The advice is simple: start with one bounded task on a copy of a file, then run the same task in the other assistant on a fresh copy and compare what you can verify.
Read next
MCP vs. Computer Use for Creative Apps: Which Should You Use?
Claude vs. Codex for 3D Modeling: Blender, Maya & 3ds Max (the same comparison for 3D)
SolidWorks with Claude and Codex
SolidWorks is the most-read pair in this group, and the two guides are deliberately parallel. Read the one for your assistant, then skim the other for the differences.
Both explain the same architecture. A community MCP server talks to SolidWorks through Windows COM or the SOLIDWORKS API. The assistant calls its tools to inspect the active document, read the feature tree, list assembly components, configurations and mates, and read custom properties. Depending on the connector, it can also create drawing views, add dimensions and balloons, build a bill of materials and export PDFs. Computer use adds the visual side: opening the right file, navigating the FeatureManager tree, switching configurations, reading rebuild warnings and checking that a drawing sheet looks right.
The shared advice is to use MCP for deterministic, repeatable operations and computer use for navigation and visual confirmation. Both guides say precise mouse work, such as dragging sketch entities, is fragile. Both also say this is not a native SolidWorks feature. You need Windows, your own SolidWorks install and license, a connector and a test copy of the file.
The practical differences are in setup. The Claude guide registers the server with a claude mcp add command in Claude Code or Claude Desktop. The Codex guide uses an [mcp_servers.solidworks] entry in Codex's configuration. Both start with a read-only request ("is this a part, assembly or drawing?") before anything that can change the model.
In the comparison post, the check for a SolidWorks edit is the feature tree and the saved part. Inspect the dimension, look for rebuild errors, confirm which configuration changed, and then reopen the saved copy. A hole's diameter surviving the edit does not mean its position did.
Read next
Best AI Tools for SolidWorks (a wider look at AI that helps with real SolidWorks work)
AutoCAD with Claude Code and Codex
AutoCAD is where the two guides diverge most, because they each pick a different connector.
The Claude Code guide compares three options. Autodesk's own AutoCAD and Civil 3D MCP Server is a public beta that runs inside AutoCAD or Civil 3D and mainly reads and analyzes drawings. The guide notes that Autodesk documents Autodesk Assistant as its supported client, so it should not be presented as a ready-made Claude Code integration. The open-source AUTOCAD-MCP project exposes tools such as list_documents, get_active_document, send_command and eval_lisp, targets AutoCAD 2026 and uses a local HTTP connection.
The Codex guide builds on the AUTOM8LABS MCP Connector for AutoCAD. It runs as a plugin inside AutoCAD, and the path is Codex, then the MCP server, then the AutoCAD plugin, then the live DWG. It supports full AutoCAD 2022 through 2027 on 64-bit Windows 10 or 11. AutoCAD LT is not supported, because LT does not run the add-ins the connector needs.
Both guides use the same safety pattern: read-only queries first (layers, blocks, layouts, entity counts), command and AutoLISP execution only when needed, and approval for anything that changes the drawing. Either assistant can also write an AutoLISP routine for you to review and load manually, with no MCP at all.
The comparison post adds a useful test: report title block attributes, find missing values, and update only approved entries. The traps are real. Attribute tags are not displayed text, and a block reference is not its block definition. Replacing a field with plain text can look right while breaking later updates.
Read next
Top AI Plugins for AutoCAD (separates built-in features from plugins and standalone tools)
CATIA with Claude and Codex
CATIA needs the most care before you start. Establish whether you are on CATIA V5 or 3DEXPERIENCE, which release, and which automation route is available. A macro, a COM client and a CAA extension are different things with different setup requirements.
Both CATIA guides describe the same community-built CATIA V5 MCP server. It connects to a running CATIA session through Windows COM, and can read the active document, inspect parts, products and the specification tree, read parameters, create supported sketches and Part Design features, and export to formats such as STEP, IGES or STL. It exposes only a subset of CATIA's automation, and it is not an official Dassault Systèmes integration.
The twin posts differ on one point that matters. The server is documented for Claude Desktop and Claude Code, so the Claude guide describes a documented setup. The Codex guide says Codex compatibility has to be tested separately. Neither promises that it works with 3DEXPERIENCE.
A good first test, from the comparison post, is a named length parameter: inspect it, propose a change, apply it on a copy, and check the dependencies and the updated model. Confidently generated V5 code for the wrong platform is not an advantage.
Read next
Claude vs. Codex for CAD (includes a CATIA parameter test)
Revit, Fusion, Rhino, SketchUp, Inventor, FreeCAD and KiCad
The same pattern repeats across the rest of the CAD and design apps: a community connector, Windows for most of them, an inspect-first workflow and a copy of the file.
Revit. A server forwards requests to a Revit add-in, which dispatches them through the Revit API. Claude can read levels, rooms, views, sheets, families and parameters. Guides exist for Claude and Codex.
Autodesk Fusion. The Claude guide separates the local Fusion MCP Server (the live session) from Autodesk's cloud-based Fusion Data MCP Server (hubs, projects and folders). There are twin guides for Claude and Codex.
Rhino and Grasshopper. A local plugin or bridge exposes layers, curves, surfaces, Breps and Grasshopper definitions. See the Claude and Codex guides.
SketchUp. The Claude guide follows one interior model through three finish options and a consistent set of exported views. A Codex version also exists.
Inventor. Automate Autodesk Inventor with Claude Code uses a third-party integration with an MCP server plus a version-specific add-in, applied to producing three length variations of a mounting plate with a STEP export for each. iLogic is a complementary option.
FreeCAD and KiCad. Both use third-party connectors. The FreeCAD guides (Claude, Codex) stress keeping the editable parameters, not just the shape. The KiCad guides (Claude, Codex) walk through a board review and manufacturing output checks, with you responsible for the electrical decisions.
Read next
Run the CAD app and the agent on one Windows machine
Most connectors in these guides use Windows COM or an add-in loaded inside the application. That means the CAD app, the connector and the assistant need to run on the same Windows host. On a Mac, or on a laptop that struggles with CAD, that is the practical blocker.
How to Run AI-Assisted Windows CAD Workflows from a Mac covers exactly this. Your Mac is only the screen you use to reach a Windows desktop. The CAD app, the AI tools and the project files all run there. It separates three ways AI can take part (a generated script, MCP, computer use), each with its own setup and verification.
A separate machine also helps with trust. The Claude and Codex security guides suggest giving the agent a computer built around the task, with your personal browser profile and unrelated client folders left off it. Both say separation is a starting point, not a guarantee: files you upload, accounts you connect and tools you authorize still need attention.
Read next
Which app, which first check
Your CAD app | Check first | Claude | Codex |
|---|---|---|---|
SolidWorks | Windows host, your license, a connector that exposes the operations you need | ||
AutoCAD | Full AutoCAD vs LT; connector support for your version | ||
CATIA | V5 or 3DEXPERIENCE; which automation route exists | ||
Revit | Windows, a Revit add-in plus MCP server | ||
Fusion | Local session server vs cloud Data server | ||
Rhino / Grasshopper | A plugin or bridge, and a duplicate .3dm file | ||
SketchUp | Connector tools, shared components and materials |
A fair test on your own files
The comparison guide lays out a test that needs no benchmark. Pick one operation from your real work: a dimension change in SolidWorks, a title block update in AutoCAD, a parameter revision in CATIA. Give each assistant its own copy of the same file, the same CAD release, the same documentation and, where possible, the same integration. Write the acceptance checklist before you run anything. Run one assistant at a time and keep the code, tool results and final files.
Then score the same things for both: completion against your checklist, correct targeting (document, object, configuration or layout), side effects, how many corrections you had to make, whether the saved file reopens correctly, and whether it works again from a fresh copy. One good run shows the workflow is feasible. It does not show reliability, so repeat it before you adopt it.
Doing this on Vagon
To let Claude or Codex work in your CAD app, the app, the connector and the agent have to share one machine. A Vagon Cloud Computer gives you that machine as a Windows or Ubuntu desktop with a GPU in a browser tab. The AI agents page recommends Spark and Planet for agents, and you can switch to a bigger performance in seconds when a model gets heavy. You pay by the minute (pricing).
Vagon doesn't ship a Claude or Codex integration for CAD, provide AI access or include licenses. You bring the connector, the agent sign-in and the CAD license. SolidWorks needs Windows. Pick your CAD app when you create the computer and check the installed version, rather than assuming SolidWorks or CATIA are there.
Setup guides: Claude Code and Codex. Start read-only, work on a copy, and review every change.
FAQ
Is Claude or Codex better for CAD automation?
There is no established winner across SolidWorks, AutoCAD and CATIA. Vagon's comparison guide says to test both on the same bounded task, with equivalent documentation and tool access, and to judge the verified file, the side effects and the corrections needed.
Can Claude or Codex control SolidWorks directly?
Only through a connector. Both can use a compatible MCP server, but the server has to expose the SolidWorks operations you need, and SolidWorks plus its license must run on Windows. A normal chat has no access to an open part.
Do I need MCP to use AI with AutoCAD?
No. Either assistant can write an AutoLISP routine that you review and run yourself. MCP adds a live connection to an open drawing, and it needs a connector that supports your AutoCAD version. AutoCAD LT is not supported by the AUTOM8LABS connector described in the Codex guide.
Does a CATIA V5 integration work with 3DEXPERIENCE?
Do not assume so. The community CATIA server in Vagon's guides targets V5 through Windows COM automation. Confirm the product, release and automation interface before you build on it.
Are CAD licenses included on a Vagon cloud computer?
No. Vagon's AI agents page says you bring your own licenses for apps such as SolidWorks, and you bring the agent too. Vagon provides the Windows computer, not the application license or the MCP connection.
Can an AI assistant replace engineering review?
No. Finishing an automation task does not show that a design is suitable, manufacturable or compliant. Qualified reviewers stay responsible for the decisions and the deliverables.
Read next
All posts in this guide
All guides →
Keep going
Related guides
All guides →



AI agents
6 chapters
AI agents for creative and CAD software: the guide
An AI agent can only work with a creative or CAD app through a connection you set up: a script, an MCP server or computer use. Pick the narrowest connection that does the job, start read-only on a copy, and give the agent its own machine to run on.
Read →



Hardware
6 chapters
Best GPU for 3D, CAD and video work
Match the GPU to the app, and look at VRAM before speed. CAD and editing work depends on drivers, stability and the CPU as much as raw GPU power. GPU renderers, real-time engines and AI features run out of VRAM first. If your needs change by project, renting by the minute can beat buying.
Read →



AI agents
6 chapters
Cloud Computers for AI Agents: How to Choose
An agent needs a cloud computer when the job involves real applications, files and a desktop you can watch, not just code. A sandbox bounds what a process can touch, a local PC is the easy start, and a GPU only helps when the work the agent triggers can use one. Run the agent inside the computer, and start with one bounded job.
Read →
Software
Software in this guide
Browse software →






















