
Cloud Computers for AI Agents: How to Choose
TL;DR
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.
Updated
·
6 chapters
·
8 min read
Start here: read the first post →
See the chapters ↓
In this guide
What you’ll learn
6 chapters

01
What changes when an agent has its own computer
The model thinks, the computer does the work, and the setup decides what you can delegate.
Jump to chapter ↓

02
Sandbox or cloud desktop
A sandbox sets boundaries. A desktop is a place to work. Which one your agent needs.
Jump to chapter ↓

03
Local PC or cloud workstation
When delegated jobs start competing with your own work, or need Windows-only software.
Jump to chapter ↓

04
Does your agent need a GPU?
Separate the model from the tools it runs, and find the real bottleneck before paying for more GPU.
Jump to chapter ↓

05
Keep it secure
A separate computer is a starting point, not a guarantee. Scope access and run the agent inside it.
Jump to chapter ↓

06
When the connection fails
A symptom-first path for MCP servers that will not connect in Claude Code or Codex.
Jump to chapter ↓
An AI agent can write a script in seconds. Finishing the job often means opening an application, loading files, running a render or build, and checking the result. That is where the question of where the agent works starts to matter. This guide is for developers, creators and teams deciding between a sandbox, their own PC, a coding environment and a cloud desktop. It covers what changes when an agent gets its own computer, when a GPU helps, and how to keep the setup contained.
What changes when an agent has its own computer
The main change is what you can delegate. Vagon's guide to choosing a setup says you can hand over a bounded sequence instead of one question: inspect the inputs, find what is missing, prepare the files, run the process, check the output. You still supply the tools and permissions, and you still decide which inputs are correct, what changes are allowed and whether the result is good.
Two points keep expectations honest. First, a cloud computer does not make the model smarter or move it onto the cloud machine. Set Up an AI Workspace on Vagon notes that prompts, selected files, screenshots and tool results may still be processed by the AI service you use. Second, the model, the agent tools and the machine have different jobs. The model reads your request, the tools give file and terminal access, and the computer is where the applications actually run.
The practical payoff is visibility. On a desktop, you can watch the agent work: it drives a real browser or app, and you can take the mouse, fix the one thing it got stuck on, and hand control back. A headless server cannot show you that.
Read next
Sandbox or cloud desktop
AI Sandbox vs. Cloud Desktop draws a clean line. A sandbox is a boundary around execution, built from operating-system restrictions, containers or isolated virtual machines. Its scope decides what it protects: a process might be allowed to change one project directory and blocked from reading anything else. A cloud desktop is a workplace. Applications, dependencies, project assets and agent tools sit in the same environment, and you can enter it yourself.
If the agent processes a spreadsheet and returns a report, isolated execution may be enough. If it revises a video project, it needs the editor, plugins, fonts and media, which is a workplace problem. The two are not exclusive: the post covers using sandbox controls inside a desktop.
Vagon vs. GitHub Codespaces covers a related fork, and it starts with the part that favors Codespaces. For pure coding in a repo, Codespaces is purpose-built and often cheaper. It has no desktop and no GPU. Vagon is a full Ubuntu desktop with an optional GPU, so it fits graphical, GPU or non-coding work. The post notes that many teams use Codespaces for the coding loop and Vagon for the rest.
Read next
Local PC or cloud workstation
Your own PC is the easiest start, and Local PC vs. Cloud Workstation for AI Agents says so. The question changes when delegated work competes with yours. A build uses CPU and memory, a render occupies the GPU, and a batch conversion reads and writes large files. If you need the same resources for another project, you either wait or both slow down.
A second workstation separates the agent's execution from your everyday device. The post names the patterns that signal it is time: you postpone agent tasks because they would slow your machine, the job needs Windows-only software, or jobs regularly exceed your device's resources. It also lets you keep editing locally, then return to the same environment to inspect the output.
The main guide adds a practical test: move one real job first and check the handoff before you move more. It also asks you to think about what remains between sessions, because saved files, running processes and a reproducible setup are three different things.
Read next
Does your agent need a GPU?
When Does Your AI Agent Workflow Need a GPU? starts by separating the model that answers requests from the tools that carry out actions. If the model is hosted online and the tools run on your workstation, the provider handles inference. A GPU in your machine does not speed up that remote model. It helps only when the work executed on your machine can use it.
Workloads that can benefit include rendering and real-time 3D (the post gives Blender Cycles as an example of GPU rendering), video and image processing, and local model inference. Even then, the post warns that not every 3D operation is GPU-bound. Scene preparation, geometry processing and simulation can depend on CPU or system memory, so find the slow step before buying more GPU. Low GPU use during an automated task is not automatically a problem either, for example when the agent is waiting between steps.
For local models, VRAM decides what fits, not just how fast it runs. Memory needs go beyond the file size because longer context uses more. If you want to run models yourself, read Run local AI models on a cloud GPU.
Read next
Keep it secure
Both security guides open with the same advice: give the agent its own computer, not your everyday desktop. Your daily machine has downloads, saved browser sessions, unrelated client folders and credentials. A dedicated cloud computer lets you build the workspace around one task. Claude Code securely on a cloud computer and Codex securely on a cloud computer are explicit about the limits.
Run the agent inside the cloud computer. Starting it locally and pointing it at a remote desktop does not move its file access or terminal permissions off your machine. Separation is not a guarantee, and files you upload, accounts you connect and tools you authorize still extend access. The Codex guide pairs the separate computer with Codex's own sandbox and approval controls.
The watch-it-work guide adds two more points. Every machine is an isolated VM, so damage from a bad plugin stays on that machine. But isolation does not stop prompt injection or vet the plugins your agent installs, so use scoped, throwaway credentials.
Read next
When the connection fails
MCP Server Not Connecting? is worth bookmarking. Its method is to identify which layer failed before you reinstall anything or raise timeouts. Did the client load the configuration, initialize the MCP connection, and reach the application? Capture the exact error and when it appears, then check the transport: stdio launches a local process, Streamable HTTP connects to an endpoint, and they fail differently. The commands use PowerShell on Windows, with WSL differences explained.
Keeping the assistant, the MCP server and the application on one desktop removes a layer: fewer connections between separate machines, and you can inspect each part in place.
Read next
Which setup to pick
Your situation | Likely fit | Why (from the posts) |
|---|---|---|
Agent processes files or data, no desktop app | Sandbox | Bounded execution is enough |
Agent works in a repo only | Codespaces or local | Purpose-built for coding, often cheaper |
Agent drives Blender, an editor or CAD | Cloud desktop | The app, plugins and media live in one place |
Agent jobs slow your own work | Cloud workstation | Separates execution from your device |
Windows-only software on a Mac | Cloud workstation | Runs the app where it is supported |
You want to watch and take over | Cloud desktop | Live desktop, instant takeover |
Rendering or local model inference | Cloud workstation with a GPU | Only when the workload can use it |
Doing this on Vagon
The easiest way to give an agent a real computer is to give it one that isn't yours. A Vagon Cloud Computer is a separate virtual machine reached over an encrypted connection. Choose Windows or Linux (Ubuntu), add the apps you need so they are installed before you connect, then install Claude Code, Codex, OpenCode or your IDE next to them. Register the app's MCP server, use computer use for visual checks, and take over whenever you want.
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. Setup is on the AI agents page and rates on pricing.
Vagon gives you the machine. The app connection depends on the MCP server you choose, and many are community-built, so start read-only. Bring your own app licenses and AI subscription.
FAQ
Why would an AI agent need a cloud computer?
Because many jobs need more than code. Vagon's guide describes inspecting inputs, installing dependencies, opening an application, running a process and checking the result. A cloud computer gives the agent a place to do those steps with approved tools, separate from your everyday desktop.
Is an AI sandbox the same as a cloud desktop?
No. A sandbox limits what covered processes can access. A cloud desktop is an interactive computer with applications and files. They can overlap, and you can use sandbox controls inside a desktop.
Does my AI agent need a GPU?
Only when the work it triggers can use one, such as GPU rendering or local model inference. A cloud-hosted model does its inference on the provider's side, so adding a GPU to your own computer does not make that model respond faster.
Should I use GitHub Codespaces or a Vagon cloud computer?
Vagon's comparison says Codespaces is purpose-built for coding in a repo and is often cheaper for that. Vagon is a full desktop with an optional GPU for graphical, GPU or non-coding work. Many teams use both.
Is running an agent on a cloud computer automatically secure?
No. The guides call separation a starting point. Files you upload, accounts you connect and tools you authorize still need attention. Isolation contains damage but does not stop prompt injection.
Are AI subscriptions or app licenses included with Vagon?
No. Vagon provides the computer. You supply the agent, its subscription, the applications and their licenses.
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 →



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
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 →















