An AI workflow can quickly become a desktop setup project. You need the assistant, the applications it will work with, supporting tools, and access to the right files. Installing everything on your everyday computer also puts experimental scripts and new permissions alongside your personal workspace.

Give your AI agent its own computer instead.

With Vagon Cloud Computer, you can prepare a separate Windows desktop with preinstalled options such as ChatGPT Codex, Claude, Blender, and DaVinci Resolve. Bring over the files a task needs, configure the required access, and use that workspace to create, inspect, edit, and export the result.

Preinstallation reduces setup work, but accounts, licenses, and AI connections may still need configuration. This guide walks you through choosing an app stack, preparing your workspace, and completing a first controlled task without rebuilding your personal computer around the experiment.

Give Your AI Agent a Computer, Not Your Everyday Desktop

Your personal computer already contains more context than most AI tasks need: client projects, saved browser sessions, personal documents, and tools configured for your daily work. A separate workspace lets you start with a narrower set of files and applications instead of placing every experiment in that environment.

Vagon gives you a Windows desktop where the AI client, creative software, and working files can live together. You access it from your own device, but the applications run on the cloud computer.

That last distinction matters. Using a cloud desktop does not make Claude or Codex an offline model. Prompts, selected files, screenshots, and tool results may still be processed by the AI service according to the features and settings you use.

Put the Agent in the Workspace You Want It to Use

Install or launch the AI client inside the Vagon desktop. Open its working folder there, and configure any application connections against the software running in that environment.

For computer use, verify that the controlled desktop is the Vagon Windows session. Running a broadly permissioned agent on your personal computer and showing it a Vagon browser tab does not establish the same separation.

Once configured appropriately, the cloud desktop becomes a place to carry out the work, not just generate instructions. A Blender script can be prepared, reviewed, executed, and checked against the resulting scene without moving between machines. Supporting tools and application settings stay with the workspace for subsequent tasks.

Separate the Environment, Then Limit Its Contents

Start with project copies and only the accounts required for the task. Avoid importing your entire personal browser profile or connecting unrelated storage simply for convenience.

This separation is useful, but it is not a security guarantee. Files, accounts, and services available inside the cloud workspace may still be affected by approved actions or mistakes. Keep permissions scoped, retain backups, and review important changes.

The practical benefit is control over what you bring into the workflow, while keeping your everyday desktop out of the initial setup.

Choose a Small App Stack Around One Real Task

Start with the result you want, then select the applications needed to produce and check it. You do not need to install every AI tool or creative application before the workspace becomes useful.

For a first task, choose one assistant and one primary production application. Add supporting tools only when the work calls for them.

These combinations are starting application sets, not promises of ready-made AI connections. Selecting Claude and Resolve, for example, does not automatically configure timeline access. The assistant may need an application-specific bridge, supported computer use, or a script-based workflow.

Choose the Assistant You Can Configure and Review

Use the assistant that fits your account access and preferred way of working. You can begin with either ChatGPT Codex or Claude without configuring both for the same first task.

Keep the distinction between products clear. The Claude preinstall option does not, by itself, confirm that Claude Code is installed. If your intended workflow specifically requires Claude Code, check and complete that setup separately.

Likewise, an assistant being installed does not establish that it has permission to control other applications.

Give the First Workspace a Concrete Purpose

For this guide, we will use Blender to create a simple product-preview scene, save an editable .blend file, and render a PNG.

Source: Blender Manual

That task gives the workspace a clear test: can the assistant help prepare the work, can Blender execute it, and can you open and inspect the saved outputs?

Blender can handle both the scene and render, so the initial setup does not need another production application. You can expand the workspace after this first loop works, rather than troubleshooting several installations and integrations at once.

Create Your Vagon and Select the Preinstalled Apps

With the first task defined, create the computer that will hold your assistant, Blender, and working files. You are preparing a reusable desktop, so choose a configuration that can support the initial project without adding resources you do not yet need.

1. Create Your Account and Choose a Region

Create a Vagon account and begin the Cloud Computer setup. Review the available regions and choose one close to where you will usually work.

Connection responsiveness matters when navigating a viewport or reviewing an agent’s interface actions. A nearby region is a useful starting point, but your internet connection and network conditions also affect the experience.

Keep this as your individual Cloud Computer setup. Team management and shared workspace features are separate considerations.

2. Match the Computer to the Application Workload

For our Blender preview task, choose a GPU-equipped option appropriate to the scene and intended rendering method. Check the available CPU, RAM, GPU memory, and storage specifications rather than choosing solely by the performance tier’s name.

The resources serve different purposes:

• CPU and RAM support application operations, scripts, and working with scene data.

• GPU resources support viewport work and compatible rendering workloads.

• Storage holds applications, project assets, generated data, and outputs.

The cloud GPU does not accelerate Claude’s or Codex’s hosted model processing. It powers supported workloads on the Vagon computer.

Source: Blender Manual

You can review Vagon’s performance options as the work develops. Start with a suitable configuration, then reassess when real scene or rendering demands justify a change.

3. Select Your AI Client and Blender

In application selection, choose ChatGPT Codex or Claude, then select Blender. These are available preinstall options in the current selection information provided for this guide.

For now, leave out applications that do not contribute to the first task. If you later need a code editor, video editor, or another production tool, you can prepare that additional workflow separately.

Remember that selecting an application requests its installation. It does not include every subscription, activate paid software, or connect the assistant to Blender automatically.

4. Start the Desktop and Check the Applications

Complete the computer creation process and open the Windows desktop. Launch your selected AI client and Blender to confirm that both open.

Check the Blender version, note any sign-in or update requirements, and verify that you have enough free storage for the working files. Make sure you are launching the applications inside Vagon, not similarly named apps on your personal device.

At this point, you have the foundation in place. Next, complete account access and choose how the assistant will work with the application.

Finish the Setup That Preinstallation Does Not Cover

Preinstalled applications save you from starting with an empty desktop. They do not remove the account, licensing, or access decisions that make the workspace usable.

Before giving the assistant a task, separate what is installed from what is connected.

Complete sign-in yourself inside the Vagon desktop. Do not paste passwords or authentication tokens into a conversation. If the application requests a subscription or license, review that separately from your Vagon plan.

Choose How the Assistant Will Perform the Work

For the first Blender task, you have three possible approaches. Choose one deliberately rather than enabling every available form of access.

Script assistance is the simplest starting point. The assistant prepares a Blender Python script, you review it, and you run it inside Blender on the cloud desktop. This does not require an MCP connection or desktop-control permission. The assistant needs feedback from you to assess errors and results it cannot otherwise inspect.

MCP-connected execution requires a compatible Blender bridge. Configure it inside the cloud environment and verify which tools it exposes before authorizing changes. A successful connection does not necessarily include scene inspection, script execution, and viewport captures; check those capabilities individually.

Computer use lets a supported assistant operate approved desktop applications visually. Both Codex and Claude Desktop document this capability, subject to their current access requirements. Enable it within the Vagon session and confirm that it targets the cloud desktop. On Windows, keep the required application visible while the task runs.

Source: OpenAI Documentation

Start with the Least Access the Task Needs

For this guide, we will begin with script assistance and manual execution. It provides a useful first result without making a third-party bridge a prerequisite. You can add connected execution once the basic workspace is working.

If you specifically want Claude Code, check its installation requirements and install it separately if necessary. The Claude preinstall option alone does not establish that Claude Code is present.

Finally, test access with a small request: ask the assistant to describe the task and prepare a plan without running commands or changing files. Establish the working boundaries before proceeding to production actions.

Bring Only the Files This Workspace Needs

A separate desktop is most useful when you make deliberate choices about what enters it. Start with the files needed for the first task, not a copy of your entire personal workspace.

Use Vagon Files to transfer selected assets and project copies from your device. You can prepare transfers even while the cloud computer is stopped, then organize the files inside Windows when you begin work.

For our Blender preview, that might mean a short brief, an approved reference image, and any model or texture you have permission to use. If you are starting from a new scene, a written brief may be enough.

Separate Inputs from Working Files and Outputs

Create a simple folder structure inside the cloud desktop:

Put the task requirements in brief, preserve supplied assets in input, and save scripts and editable projects in working. Use output for rendered images and other deliverables.

This organization makes instructions clearer. You can tell the assistant where to find references, where to prepare a script, and where the final render belongs.

Folder names are not access controls. Calling a folder input does not prevent an agent or application from changing it. Use project copies, appropriate permissions where available, and explicit task boundaries. Keep the original files and a separate backup outside the experiment.

Preserve Dependencies Without Importing Everything

Existing projects may depend on textures, linked assets, fonts, or other external files. Include what the application needs to open the project correctly, then check for missing dependencies before asking the assistant to work.

Avoid transferring unrelated documents, saved credentials, or your full browser profile. Sign in only to services required for the task, remembering that connected accounts can expose more than the files visible in your working folder.

If an integration requires an API key, use its supported credential configuration rather than placing the key in the brief or a script shared with the assistant. Also check project copies for embedded secrets or sensitive material.

Before continuing, open the transferred assets yourself. Confirm that the cloud workspace contains the intended copies and that the output location is ready.

Complete a First Task and Verify the Output

Your first task should prove that the workspace can produce something useful without requiring a complicated integration. For this walkthrough, the assistant will prepare a Blender Python script. You will review and run it in Blender, then inspect the editable scene and rendered image.

This is script assistance, not autonomous application control. The assistant cannot assume it has seen the result merely because it wrote the code.

Give the Assistant a Bounded Brief

Open your selected AI client inside Vagon and describe the task. Use the actual Windows path to your working folder in place of the placeholder below.

Prepare a Blender Python script for a simple product-preview scene.

Create a cylindrical container with a separate lid, a ground plane,
a camera, and lighting. Use simple materials and no external downloads.

Workspace: [full Windows path to AI-Workspace]

Prepare a Blender Python script for a simple product-preview scene.

Create a cylindrical container with a separate lid, a ground plane,
a camera, and lighting. Use simple materials and no external downloads.

Workspace: [full Windows path to AI-Workspace]

The named collection and output paths make the result easier to inspect. They are instructions, not technical restrictions on what a script could do.

If your chat interface cannot write files into the workspace, ask it to return the script as text. Save it through Blender’s Text Editor or a code editor inside the cloud desktop.

Source: Blender Manual

Review Before Running

Check that the script follows the brief. Look for unexpected downloads, package installation, network requests, broad deletion commands, or writes outside the intended folders. Confirm that it uses the installed Blender version appropriately and handles existing output files.

Source: Blender Manual

Ask the assistant to explain anything you do not understand. Its explanation is helpful context, but not a guarantee that the code is safe or correct.

For this first run, open a fresh Blender scene without valuable unsaved work. Keep other project files closed, review the script in the Text Editor, and run it only when you are comfortable proceeding.

Check the Saved Files, Not Just the Success Message

After execution, inspect both Blender and the workspace folders:

• Confirm that AI_Preview contains the expected objects.

• Check that the lid is a separate object and the scene remains editable.

• Verify that the .blend file and PNG exist at the intended paths.

• Open the rendered PNG and inspect its framing, lighting, and visible geometry.

• Reopen the saved .blend file to confirm it contains the expected scene.

Source: Blender Manual

If execution fails, provide the relevant error message and script section to the assistant. Do not grant broader access simply to bypass a problem.

Finish with one small revision, such as changing the container’s color while preserving the camera and geometry. Save it under a new filename. A successful revision demonstrates a usable working loop: prepare, review, execute, inspect, and improve, all within the Vagon workspace.

Keep the Separation Useful as You Add More Tools

Once the first task works, you can add applications, MCP bridges, or computer use where they serve a clear purpose. Keep the same deliberate approach: each addition should solve a specific problem, not simply give the assistant more access.

For example, a Blender bridge might let the assistant inspect scene objects and receive execution errors directly. Computer use could help it check a dialog or inspect a panel that the bridge does not expose. Neither requires connecting unrelated accounts or making every project available.

Review New Connections Before Enabling Them

Check who maintains an MCP server or skill, what it installs, and which operations it permits. A connection that can execute arbitrary code has a different scope from one limited to reading scene information.

Use the assistant’s available permission controls and review requests for additional access. A project folder selection or an app approval should not be treated as a universal boundary across shell commands, connected services, and desktop actions.

Source: OpenAI Documentation

Files, websites, and other content encountered during a task can contain misleading instructions. Prompt injection remains a concern inside a cloud workspace. The separate desktop limits what you initially bring into the environment; it does not make all content trustworthy.

Preserve a Reviewable Working Loop

Keep original assets and backups separate from experimental changes. Ask the assistant to identify affected files before substantial edits, and require confirmation before deleting important data, overwriting originals, or sharing outputs externally.

After a task, inspect the deliverables and retain the editable source files. Remove connections you no longer need, sign out of unnecessary accounts, and review saved permissions before starting a different project.

Source: Blender Manual

Vagon is a persistent workspace, not an environment that automatically becomes clean after every session. Treat the files, credentials, and tools left inside it as part of the next task’s starting conditions. That keeps the workspace useful as your workflow grows.

Understand the Cost Before Expanding the Workspace

An AI workspace has several cost components. Selecting preinstalled applications simplifies setup, but it does not bundle every service into one price.

Review the current Vagon pricing for your configuration before starting. Do not assume that stopping the computer removes recurring storage or plan charges.

Begin with the Resources Your First Task Needs

Use the preview task to assess application responsiveness, memory needs, and rendering time. You do not need to provision for an entire production pipeline before completing that first test.

Vagon’s performance options let you change resources while retaining your desktop, applications, and files. Save your work and finish active operations before switching. Increase resources when the application workload justifies it, not because an assistant suggests that more GPU power will improve its reasoning.

Check the Computer State When You Finish

Before ending a session, confirm that renders and exports have completed, save the editable project, and transfer or back up important outputs. Then use Vagon’s controls to stop the computer when you no longer need it running.

Closing the browser window is not a reliable substitute for checking the computer’s actual state. Make that final check part of the workflow, alongside reviewing the files you produced.

Final Thoughts

You do not need a complicated automation stack to give your AI agent its own computer. Start with a separate desktop, one assistant, one production application, and a task you can review.

Create your Vagon Cloud Computer, select ChatGPT Codex or Claude alongside Blender, and bring over only the files you need. Complete sign-in and permissions, then prepare, run, and inspect your first output.

Once that loop works, expand the workspace around real production needs. You will have a place to develop the workflow without rebuilding your everyday computer around each experiment.

FAQs

Are Claude and ChatGPT Codex available as preinstalled apps?

Yes. Both appear in Vagon’s application selection. Choose one alongside the production applications you need, then complete sign-in and check which features your AI account provides.

Does selecting Claude also install Claude Code?

Do not assume so. The Claude preinstall option alone does not confirm that Claude Code is installed. If your workflow specifically requires Claude Code, check the cloud desktop and complete its setup separately.

Are AI subscriptions and software licenses included?

Preinstallation does not automatically include paid subscriptions or licenses. Your AI account, production software, plugins, and external services may have separate requirements and costs. Blender itself is free to use.

Do preinstalled apps come with MCP connections?

Not automatically. Application installation and AI integration are separate steps. An MCP workflow requires a compatible server or bridge, configuration, and appropriate permissions. Script assistance can be a simpler starting point.

Can the agent access my personal computer?

An agent running inside Vagon does not automatically gain access to your personal computer. However, shared folders, connected storage, accounts, and other integrations can expand what it can reach. Verify where the agent runs and which desktop computer use controls.

Do I need a GPU-equipped cloud computer?

Choose resources around the application workload. Blender rendering, complex 3D scenes, and some video tasks benefit from suitable GPU resources. File preparation and lighter scripting may have different needs. Hosted AI model processing does not run on Vagon’s GPU.

What happens to my files and costs between sessions?

Vagon provides storage for keeping applications and files between sessions, subject to your plan. Running-computer usage and recurring plan or storage charges are separate. Save and back up important work, and check that the computer is stopped when you finish.