A material change can be easy to describe and surprisingly awkward to make from a tablet. You want a darker finish, the same lighting, and two comparison renders, but you still need access to the Blender scene and a way to check what changed.

With Vagon Cloud Computer, you can use your iPad to access a desktop where Blender, Claude, Codex, and your project files work together. The cloud computer runs Blender and handles rendering; your iPad becomes the interface for giving instructions, inspecting the scene, and reviewing results.

You do not need both assistants or every integration to begin. A reviewed Python script may be enough for one task, while MCP or computer use can help with another.

This guide walks through building that workspace, connecting an assistant, and producing checked material variants without turning your iPad into the machine doing the rendering.

Source: Apple Newsroom

What Runs on the iPad, and What Runs in the Cloud?

In this setup, your iPad is the device you use to reach the workspace. Blender runs on the Vagon computer, alongside the AI client, integration components, and project files.

That arrangement keeps the production tools together. When an assistant prepares a script or requests a supported Blender operation, the relevant software is already in the environment where the scene is open. You can then inspect the result through the same remote desktop.

Opening Claude or ChatGPT on your iPad does not automatically connect it to Blender on another computer. Mobile access to an assistant and access to a Blender scene are separate capabilities. For this guide, use the AI client inside the cloud desktop so its local tools can work alongside Blender.

You can begin with Claude or Codex. Installing both gives you options, but there is no need to have two agents changing the same scene. Use one active assistant for a task, save and inspect its work, then switch deliberately if needed.

The practical advantage is continuity: the scene, scripts, materials, and renders stay in one working environment while you use the iPad to direct and review the work.

Source: MCP for Blender

Set Up Blender and Your AI Tools on Vagon

Start with a workspace that can complete the whole task: open the scene, prepare or execute the change, inspect it in Blender, and render the result. You should not need to transfer unfinished files between computers simply to check whether an instruction worked.

1. Choose Performance for Your Blender Workload

Create your Vagon account and choose a performance option suited to the scene.

Viewport interaction, material previews, simulations, and final rendering place different demands on the computer. Consider memory, CPU performance, and GPU resources together rather than selecting hardware based only on the fact that you are using AI.

For a hosted Claude or Codex session, the local GPU is primarily relevant to workloads such as Blender rendering, not the conversation itself. Vagon’s scalable performance lets you change resources while retaining your desktop, applications, and files. Save your work and stop active jobs before changing configurations.

2. Select Blender and an AI Client

During Application Selection, choose Blender and your preferred AI application: Claude, ChatGPT Codex, or both. Add Visual Studio Code if you want a separate editor for reviewing scripts.

Source: Anthropic

Keep the initial setup focused. External asset services, additional render engines, and multiple integrations are not required for a simple material-variant task.

Preinstallation is a starting point, not confirmation that every connection is ready. Selecting Claude does not establish that Claude Code is installed, and selecting Codex does not automatically install a Blender MCP server. AI subscriptions, authentication, and integration dependencies need separate checks.

3. Open the Desktop from Your iPad

Connect to your Vagon computer through its supported iPad access and open Blender inside the Windows desktop.

Then open your chosen AI client there and complete authentication. Keep the client’s execution environment on the Vagon computer rather than accidentally starting a task in a separate hosted environment.

Do not import your everyday browser profile or unrelated credentials just to reproduce a familiar setup. Sign into the services this workflow actually needs.

4. Verify the Basics Before Connecting an Agent

Open a small scene and check that you can select objects, navigate the viewport, change a value, save a copy, and produce a preview render.

Note the Blender version. Generated Python and community add-ons may depend on version-specific APIs or installation procedures.

Finally, create a clearly named project folder for the scene, scripts, and outputs. Keep an untouched source copy separately. Establishing this baseline first makes it easier to distinguish an integration problem from a missing texture, an input issue, or an ordinary Blender configuration error.

Make Blender Comfortable to Control from an iPad

AI can reduce repetitive work, but you still need to select objects, inspect settings, and review results. Spend a few minutes making those interactions comfortable before connecting an agent.

A physical keyboard and mouse or trackpad are useful for sustained work. Touch can handle simpler interactions, but Blender’s desktop interface includes small controls and shortcuts that are easier to manage with dedicated input devices.

Source: Apple Newsroom

Test Navigation Before Changing the Scene

Open a sample scene and check that you can:

• Select an object and open its context menu.

• Orbit, pan, and zoom the viewport.

• Enter a precise value in a properties field.

• Move between Blender and the AI client.

• Save a copy using the file browser.

Test the actual combination of iPad, accessories, and remote-access method you intend to use. A shortcut that works on a directly connected Windows keyboard may be intercepted or translated differently along that path.

Blender’s input preferences offer alternatives. Emulate Numpad can help when your keyboard lacks a numeric keypad. Emulate 3 Button Mouse provides an alternative to middle-mouse input, although it changes some shortcut behavior. Enable these only after checking which controls you need.

Source: Blender Manual

Keep the Working Layout Simple

Give the viewport enough room to inspect the scene. Use the Outliner to select named objects, and switch to the relevant Properties tab when checking materials or render settings. You do not need every editor visible simultaneously.

Clear object names also help the assistant target the right item. “Change the material on Product_Body” is less ambiguous than “change the object on the left.”

Source: Blender Manual

Apple Pencil may be useful for pointing, depending on your setup, but do not assume that pressure sensitivity or sculpting behavior matches a directly connected drawing device. Test those functions separately.

The aim is not to eliminate manual control. It is to make your checks quick enough that reviewing an AI-generated change remains a normal part of the workflow.

Connect Claude or Codex to Blender Through MCP

MCP gives an assistant access to defined tools, but the connection needs components on both sides. Installing Claude or Codex alone does not make your Blender scene available to it.

One community option is MCP for Blender, which documents setups for both clients. It is not an official Blender Foundation integration. Review its current installation instructions, compatibility requirements, and security notes before using it.

Install the Blender Side of the Connection

The project uses a Blender add-on alongside an MCP server that the AI client launches. The add-on receives operations inside Blender; the server exposes supported tools to the assistant.

Install the components inside the Vagon computer, not on the iPad. Follow the project’s current procedure to install and enable the add-on, then check its connection controls in Blender’s sidebar.

Source: MCP for Blender

Keep the server and add-on versions compatible. Avoid combining an older tutorial’s package name or configuration with a newer installation.

A localhost address refers to the computer where the component runs. Keeping Blender, the server, and the AI client together avoids needing to expose the Blender connection publicly. Opening a port to the internet is not a standard requirement for this setup.

Connect Claude

Use the project’s instructions for your actual Claude client. Claude Desktop and Claude Code have different configuration paths; do not assume that changing one automatically configures the other.

After adding the server, restart or refresh the client as instructed and confirm that its Blender tools are available. If the client cannot launch a required executable, check its configured path inside Windows rather than repeatedly reinstalling the Blender add-on.

Connect Codex

Use the repository’s Codex-specific setup. Its documented options include a plain MCP connection and a plugin-based integration, with different installation requirements.

Choose one route initially. Do not assume that the standalone Codex CLI is installed merely because the desktop application opens.

For this workflow, keep only one assistant connected and acting on the scene at a time. The project warns against running multiple MCP server instances simultaneously. Save the scene and stop the previous connection before switching clients.

Verify the Connection Without Editing

Start with a small request:

Compare the response with Blender’s Outliner and scene settings. A connected status alone does not confirm that the assistant is inspecting the intended Blender instance.

Finally, treat this integration as executable access, not just chat. The project includes a tool capable of running arbitrary Python inside Blender. A request to “only inspect” defines your intent but does not create a technical read-only boundary. Save a working copy before allowing changes.

First Project: Create Two Material Variants Without Rebuilding the Scene

Use a small product scene with an existing camera and lighting setup. The task is to create two finish options while preserving the composition, making it easier to judge the material change rather than unrelated scene edits.

Choose one assistant for the task. Either Claude or Codex can receive the same instructions through a compatible connection; you do not need both working simultaneously.

Identify the Target Before Editing

Save a working copy, then ask the assistant to inspect the scene and identify the product mesh, assigned materials, active camera, and render settings.

Check those names in Blender. If the product contains several meshes or material slots, specify which surfaces should change. Also check whether the material is shared with other objects: modifying a shared material can affect more than the intended product.

Source: Blender Manual

Use a bounded request such as:

Replace Product_Body with the verified object name. “Matte” and “satin” are starting descriptions, not precise specifications, so review the proposed material values before approving them.

Source: Blender Manual

Build One Variant at a Time

Approve the first finish, inspect it in Blender, and save it under a distinct filename such as product-finish-A.blend. Produce a low-resolution preview before committing to a final render.

For the second finish, reopen the unchanged baseline and apply the second plan. Save it as product-finish-B.blend. Starting both variants from the same baseline reduces accidental differences between them.

Keep the preview settings identical across both files. Camera framing, exposure, lighting, render engine, and color management should remain consistent for a useful comparison.

Compare the Scene and the Render

Review each preview on the iPad, then check the underlying Blender settings. Confirm that the intended surface changed and that nearby objects, labels, or other material slots stayed intact.

If the finish needs adjustment, describe one change at a time, such as increasing roughness slightly while preserving the base color. Avoid asking the assistant to “improve everything,” which broadens the task unnecessarily.

The deliverable is two saved scene variants and comparable previews. A tool reporting success is not enough; verify both the visible result and the scene properties you intended to preserve.

Use Computer Use for the Parts That Need Visual Inspection

MCP is useful for structured scene operations, but some checks are easier to understand through Blender’s interface. Computer use can help an assistant open a panel, inspect a camera view, or navigate an output dialog when the chosen tool supports those interactions.

It is a separate capability, not something automatically enabled by connecting Blender through MCP.

Configure the Assistant You Are Using

For Claude, follow the current Claude Desktop computer-use instructions. Availability depends on the supported platform and plan. Do not assume that Claude Code in a terminal has the same computer-use capabilities as the desktop application.

For Codex, follow the Computer Use setup and review application access before approving Blender. Its app approvals are separate from permissions governing local commands.

In either case, the desktop-control tool should operate inside the Vagon environment where Blender is running. Your iPad displays and controls that workspace; its mobile AI applications do not automatically receive the same access.

Give It a Specific Visual Check

After generating the material variants, try a narrow request:

Compare the answer with what you can see. For the product scene, useful follow-up checks include confirming the correct file is open, inspecting the selected view layer, and looking for visibly missing textures.

Source: Blender Manual

Keep visual checks proportional to what they can establish. A screenshot may reveal an unexpected appearance, but it does not prove that every shader value, modifier, or hidden object is correct.

Avoid Competing for Desktop Control

On Windows, expect foreground interaction while the agent works. Do not simultaneously drag objects, type shortcuts, or change panels in the same cloud session. Stop or pause the task before taking over manually.

Modal dialogs, custom controls, and session interruptions can disrupt automation. If the assistant becomes uncertain, inspect the current state before asking it to continue. Repeating an action blindly can create a second change rather than completing the first.

Use computer use where the interface adds information. Keep repeatable scene edits in scripts or structured tools when those provide a clearer, more inspectable operation.

When a Reviewed Python Script Is the Better Route

You do not need an MCP connection for every Blender task. If the change is small, clearly specified, and easy to inspect, asking Claude or Codex to prepare a Python script can be a simpler approach.

For the material example, you could request a bpy script that creates a material copy and assigns it to one verified object. The assistant generates the code; you decide when to run it inside Blender.

Include the Blender version, exact object name, and material slot in the request. Ask the assistant to stop with a clear error if the target is missing or ambiguous rather than choosing another object automatically.

Before execution, check that the script:

• Uses explicit object references rather than relying unnecessarily on the current selection.

• Preserves unrelated materials, geometry, and scene settings.

• Contains no unexpected downloads, subprocesses, or file operations.

• Does not overwrite the source scene.

Run the reviewed code through Blender’s Text Editor on a working copy. Keep an independent baseline; do not assume every script operation can be safely reversed with Undo.

If execution fails, inspect the scene before trying again. A script can complete some changes before encountering an error, so rerunning it may duplicate materials or apply another modification.

Script assistance is also useful when you want to preserve the exact operation for later review. It does not give the assistant live knowledge of the scene unless you provide that information, and it does not make generated code inherently safe. The advantage is a concrete implementation you can inspect before it acts.

Save, Render, and Bring the Results Back to Your iPad

Once both finishes look right, turn the previews into deliverables. Keep the saved scene variants and rendered images clearly associated, for example:

Check the active filename before saving. The assistant may have changed the scene successfully without saving those changes to the file you intend to keep.

Source: Blender Manual

Check Dependencies and Render Settings

Confirm that textures, linked libraries, and other required resources resolve correctly. Blender’s Pack Resources can include eligible external resources in the .blend file, but not every dependency can be packed. Review external caches and media separately, and save the file after packing.

Before the final render, verify:

• The intended camera, scene, and view layer.

• Output resolution, file format, and destination.

• Render engine, sampling, and color management.

• GPU device selection, if the engine and hardware support it.

Choosing a GPU-equipped cloud computer does not by itself confirm that Blender is using that GPU for rendering. Inspect the relevant preferences and render settings.

Source: Blender Manual

Render one final image first and open the saved output. Check framing, material appearance, transparency, and any labels or fine details before rendering additional views or frames.

Transfer the Checked Outputs

Use Vagon Files to download the deliverables to your iPad’s available file location. Open the downloaded image and confirm that you received the final version, not an earlier preview. If you need the editable project elsewhere, include its required dependencies too.

Treat Disconnection and Shutdown Differently

The render runs on the cloud computer, but disconnecting your iPad, locking it, and stopping Vagon are different events. Save before leaving, keep the computer running for active render jobs, and verify your session behavior with a short test before relying on a long unattended job.

Computer-use tasks have additional active-desktop requirements and may pause for approvals or dialogs. Do not assume that a render continuing means an interactive agent can also continue.

Once the outputs are saved and checked, stop unnecessary processes and the cloud computer when you no longer need them running.

Final Thoughts

Your iPad can be the control surface for a complete Blender workflow, with scene editing, AI tools, and rendering running together on Vagon Cloud Computer.

Start with one assistant and a small scene copy. Use MCP for supported scene operations, reviewed Python for inspectable changes, and computer use when navigating the interface adds value. Check the result in Blender before committing to final renders.

Create your workspace, select Blender and your preferred AI client, and produce one verified material variant before expanding the workflow.

FAQs

Is Blender running on my iPad or on Vagon?

In this workflow, Blender runs on the Vagon cloud computer. Your iPad displays the remote desktop and sends your input. Scene processing and rendering use the cloud computer’s resources, not the iPad’s GPU.

Do I need both Claude and Codex?

No. Start with either assistant. Both can help generate scripts, and the community MCP project discussed here documents connections for both. Using two assistants is optional, not a requirement for completing the workflow.

Can the iPad apps control Blender automatically?

Not simply by opening them. An assistant needs an explicitly configured connection to the computer running Blender. This guide keeps the AI client and Blender together inside Vagon rather than assuming a mobile conversation has desktop access.

Do I need MCP, or can I use Python scripts?

You can use reviewed Python scripts without MCP. Ask the assistant to generate code for your Blender version, inspect it, and execute it manually in Blender. MCP is useful when supported tools provide scene access or operations you want to delegate.

Can Claude and Codex edit the same scene simultaneously?

Avoid that arrangement. Overlapping operations can create conflicting or difficult-to-trace changes, and the discussed MCP project warns against multiple server instances. Save, inspect, and stop one connection before switching assistants.

Do I need a keyboard and mouse?

They are useful for sustained work, particularly shortcuts, precise selection, and script review. Touch may be sufficient for simpler interactions. Test viewport navigation, right-clicking, and modifier keys with your actual access setup.

Can I use Apple Pencil for Blender?

It may be useful for pointing, depending on your setup. Do not assume pressure sensitivity or sculpting behavior matches a directly connected drawing device. Test those functions before relying on them.

Will rendering continue when I disconnect?

Disconnecting the iPad is different from stopping the cloud computer. A render may continue while the computer remains running, but verify session behavior with a short test. Interactive agent tasks can have additional desktop and approval requirements.

Are AI subscriptions included with Vagon?

Do not assume so. Preinstalled applications, AI account access, cloud compute, storage, and paid third-party services are separate considerations. Check the requirements and charges for each service you use.