
AI agents for creative and CAD software: the guide
TL;DR
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.
Updated
·
6 chapters
·
10 min read
Start here: read the first post →
See the chapters ↓
In this guide
What you’ll learn
6 chapters

01
How an agent reaches your software
Scripts, MCP and computer use are different tools, and choosing between them is the first decision.
Jump to chapter ↓

02
CAD and engineering
SolidWorks, AutoCAD, CATIA and the rest, with Claude and Codex side by side.
Jump to chapter ↓

03
3D, VFX and game engines
Blender, Maya, Houdini, Cinema 4D, Nuke, Unreal, Unity and Godot.
Jump to chapter ↓

04
Video and design
Premiere Pro, DaVinci Resolve, After Effects, Photoshop and Figma.
Jump to chapter ↓

05
GIS and engineering analysis
QGIS, HEC-RAS, SWMM, SNAP and other technical tools, driven by an agent.
Jump to chapter ↓

06
Where the agent runs
Local PC or cloud workstation, whether you need a GPU, and how to keep the setup contained.
Jump to chapter ↓
AI coding agents such as Claude Code and Codex can now work with design, 3D, video and engineering software, not just code. But an agent is not built into Blender or SolidWorks. It reaches the app through a connection you set up, and it runs on a computer that has to be able to run the app. This guide is the map: how agents connect, what Vagon has written for each family of software, and where to run it all. It is for creatives and engineers who already use the software and want to know what is realistic.
How an agent reaches your software
There are three ways an agent can work with an application, and mixing them up causes most of the confusion. A script or command-line tool does a defined job. An MCP server exposes tools that read or change application data. Computer use lets the agent look at the screen and click, type and navigate like a person.
MCP vs. Computer Use for Creative Apps gives the order to try them in. Use an existing script or command-line tool if it already does the job. Use an MCP tool when the agent needs application data or supported actions. Use computer use when a step depends on the interface, such as an export dialog, a warning or whether a title fits the frame. Combine them only where one leaves a gap. The same guide makes the point that matters most: choose how you will verify the result at the same time as how you will execute the task. A successful tool response or a completed click does not show that the saved model, timeline or export is right.
Every app guide on Vagon follows the same loop, as the AI agents page puts it: inspect, propose, approve, edit, verify. Start with a read-only request on a copy of the file. The guides also share a limit: Vagon does not ship a Claude or Codex connector for these apps. Many MCP servers are community-built, and computer use is for checking rather than precision work such as sculpting.
If your connection fails, MCP Server Not Connecting? Fix Claude Code & Codex Errors is the troubleshooting guide. For the question of whether the agent should work in a restricted sandbox or on a full desktop, read AI Sandbox vs. Cloud Desktop: a sandbox defines boundaries, a desktop provides a workplace, and the two can be combined.
Read next
CAD and engineering
For SolidWorks, a community MCP server connects through Windows COM or the SOLIDWORKS API. The agent can read the feature tree, assembly components, configurations and custom properties, and with some connectors build drawing views and export PDFs. Computer use then opens the sheet and checks the layout. See the Claude guide and the Codex guide.
AutoCAD is different because connectors differ. The Claude Code guide compares Autodesk's own beta MCP server, which is mainly read-only, with an open-source option. The Codex guide uses a plugin-based connector for full AutoCAD on Windows. CATIA rests on a community V5 server, and the Codex guide for CATIA says Codex compatibility must be tested separately.
Claude and Codex are separate intents here, and the library treats them that way. The comparison guide says neither is a universal winner and that the integration matters more than the model. The full side-by-side lives in Claude and Codex for CAD.
Read next
3D, VFX and game engines
For 3D and VFX, the pattern is the same with different objects. In Blender, a connector lets Claude inspect objects, collections, materials, modifiers, cameras and Geometry Nodes, and computer use handles the Scripting workspace and viewport checks. You need a compatible Blender version, an MCP connector, a Blender add-on and a supported Claude client, and you should review changes before saving the .blend file.
Houdini work centers on the node graph: tracing connections, diagnosing a stale File Cache and making controlled variations while Houdini does the cooking and simulation. In Unreal Engine, the guide covers Epic's experimental Unreal MCP in UE 5.8 and says other engine versions may need a third-party option. Nuke is covered as a tested Python tool, a Read-path remapper, rather than an unattended agent.
For direct comparisons, read Claude vs. Codex for 3D Modeling and Claude vs. Codex for Game Development. The first says there is no defensible universal winner from feature documentation alone, and judges results on editable geometry rather than a polished render. More in Claude and Codex for 3D and VFX.
Read next
Video and design
In video, the guides show a similar split. For Premiere Pro, Claude Code does not open a .prproj file like a text document. It talks to a running Premiere session through a local bridge, plugin or extension, and may inspect sequences, list clips, create markers and prepare captions. The matching Codex guide covers the same app. DaVinci Resolve is notable: the guide says Resolve Studio 21.1 introduced built-in integration with AI assistants through MCP.
How to Edit Videos with Claude and Codex on Vagon follows a workflow from raw media to a reviewed cut, captions and delivery versions, with editorial decisions staying yours.
Photoshop is the clearest case of "no automatic connection." The Claude Code guide says the most reliable route is to have Claude write or debug an Adobe UXP script or plugin, run it in Photoshop and return the result, with community MCP bridges as an optional, third-party extra. See Claude Code with Photoshop and Codex with Photoshop. More in Claude and Codex for video and design.
Read next
GIS and engineering analysis
The same approach reaches technical tools outside the design world, mostly through scripted, repeatable jobs. In the QGIS guides, the assistant prepares a processing script or batch commands, and QGIS performs the spatial analysis: one example finds building footprints inside a 50-metre buffer around selected roads, and the Codex guide batches reprojection and clipping with qgis_process. For hydraulics, the assistant reads results and builds comparisons while you keep the engineering judgment: HEC-RAS results live in HDF5 files, and the SWMM guide compares drainage scenarios. ESA SNAP uses its Graph Processing Tool for batch image processing.
Read them for the pattern: define the scope, run on copies, and verify the output files in the application.
Read next
Where the agent runs
The model and the workstation do different jobs. Local PC vs. Cloud Workstation for AI Agents makes the point that a hosted model can answer from afar while the files, commands and apps still run on your PC. The model interprets the request; the workstation supplies the applications, storage and processing.
That decides whether you need a GPU. When Does Your AI Agent Workflow Need a GPU? says a GPU helps when the application the agent drives can use it, or when you run a model on the same machine. An online model's answers do not get faster because your PC has a bigger graphics card. For running models themselves, see Run local AI models on a cloud GPU.
A separate computer also changes how you supervise. Cloud Computers for AI Agents frames it as separating execution from your everyday device, and How to Watch Your AI Agent Work on a Cloud Desktop explains why a live desktop beats a headless server: you can see the agent work and take over when it gets stuck. The security guides for Claude Code and Codex agree that separation is a starting point, not a guarantee. See Cloud computers for AI agents for the full path.
Read next
Choosing your approach
What the agent has to do | Start with | Then verify |
|---|---|---|
Repeat a defined job on files | An existing script or command-line tool | The output files |
Read or change data in an open app | An MCP server for that app | Object, scope and units in the app |
Handle a dialog, warning or visual check | Computer use | What the warning means; data beyond appearance |
Review a drawing or timeline visually | Computer use or captured views | Application data, not just the picture |
Processing that needs a heavy app | A cloud computer with the app installed | The saved file, reopened |
Source for the table: Vagon's MCP vs. Computer Use and CAD comparison guides.
Doing this on Vagon
An agent is only as useful as the machine it drives. It needs your creative apps installed, enough GPU to open real scenes, and a place where a mistake can't touch your own computer. A Vagon Cloud Computer is that place. Create a Windows or Linux (Ubuntu) computer with your apps preinstalled, install your agent next to them, and watch it work in the browser. If it gets stuck, take over the mouse.
Start on Spark or Planet and switch to a bigger performance when a scene or timeline gets heavy. Shut the computer down and billing stops. The AI agents page has setup guides for Claude Code, Codex, Cursor and OpenCode.
Vagon provides the machine, not the connector, AI access or app licenses. Isolation limits damage but doesn't stop prompt injection, so use scoped, throwaway credentials. For a chat agent parked around the clock, a small always-on server costs less.
FAQ
Can Claude or Codex control Blender, SolidWorks or Premiere Pro?
Not on their own. Vagon's guides say there is no native one-click connection. You add a compatible MCP server or bridge, which is often community-built, and you can add computer use for visual checks. What the agent can do depends on that connector, so start with a read-only request.
What is the difference between MCP and computer use?
MCP gives the agent structured tools that read or change application data. Computer use lets it operate the visible interface through screenshots, clicks and typing. Vagon's guide says to use a script or command-line tool first if one already does the job, then MCP, then computer use for interface-only steps.
Do I need a GPU to run an AI agent?
Not for the agent itself. A hosted model does its thinking elsewhere. A GPU helps when the application the agent drives, such as a renderer, can use it, or when you run a model on the same machine.
Does Vagon provide the AI, API keys or application licenses?
No. Vagon's AI agents page says you bring the agent, its sign-in and the application licenses. Vagon provides the computer: a Windows or Ubuntu cloud computer with a GPU and your apps, in a browser.
Is it safe to let an agent run on a cloud computer?
A separate machine contains the damage, but Vagon's guides say it is not immunity. It does not stop prompt injection or vet the plugins an agent installs. Use scoped, throwaway credentials, keep your main keys off the machine, and approve edits.
Where should I start?
Pick one app you already use and one small read-only task, such as asking the agent to report what is in the open document. Use a copy of the file. Then move to a small, approved change and check the saved result.
Read next
All posts in this guide
All guides →
Keep going
Related guides
All guides →



AI agents
6 chapters
Claude and Codex for CAD: SolidWorks, AutoCAD, CATIA
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.
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 →



AI agents
6 chapters
Claude and Codex for 3D, VFX and Game Engines
Neither Claude nor Codex wins across 3D and VFX apps on its own. The result depends on what connects the assistant to the app (a script, an MCP server or computer use) and how you check the saved scene. Start with a read-only audit of a copy, then approve one change at a time.
Read →
Software
Software in this guide
Browse software →
















