Instant Connection for Pixel Streaming

— New Feature Automated Setup

Comparison

Claude Code on the Web vs a Cloud Computer: What Runs Where

Compare Claude Code on the web with a cloud computer. See limits for desktop apps, GPUs, and MCP, and when to run your application workflow on Vagon.

On this page

Claude Code on the web can fix a bug, install dependencies, and run tests without using your computer to execute the work. If the result you need is a reviewed code change, its hosted environment may already provide everything the task requires.

An application project raises a different question. Can that environment run the software, reach the project, and expose the state Claude needs to inspect? A script for a CAD model still needs a compatible CAD runtime. A GPU render needs suitable hardware. Reviewing either result may also require opening the application yourself.

At Vagon, we provide a cloud computer where Claude Code and your applications can work together, with a desktop you can enter to inspect and continue the project. That gives application work its own execution machine while your everyday computer stays available.

The choice starts with what must run and how you need to review it. Here is where Anthropic's hosted sessions fit, where their limits matter, and when to move the task to a full cloud computer.

Claude Code on the Web Already Runs Real Development Work

The standard Claude Code web workflow starts with a repository and a task. Claude clones the selected GitHub repository into an Anthropic-managed environment, works on the code, and pushes a branch you can review. You can send follow-up instructions during the session or leave it working and return later.

Claude Code web interface showing multiple sessions and a task updating a GitHub README with file-reading and shell output.

Source: Simon Willison

This suits jobs with a clear result in the repository: correcting a function, adding tests, updating documentation, or tracing how a feature works. Independent tasks can run in separate sessions and branches. You still need to review the changes before merging them.

It Can Execute the Code It Changes

Anthropic's cloud environment documentation lists Python, Node.js, compilers, Docker, databases, and Chrome for Testing among the available tools. Claude can run commands, install additional dependencies, and start services required by a project. A test that needs PostgreSQL is therefore not automatically a reason to move to another computer.

Setup scripts can prepare dependencies before Claude starts, and environment caching can reduce repeated installation. Network access is configurable, so package downloads and external requests need an environment that permits the required destinations.

MCP tools can also be part of the setup. Supported project-level definitions and enabled connectors can make tools available, but configuration stored only on your own computer does not automatically transfer. The server, its dependencies, and the service it calls must be accessible from the session.

Give the Task a Verifiable Finish

Suppose a Python importer drops rows when an optional field is empty. A useful task would name the sample input, explain which rows should survive, and ask for a regression test alongside the correction. Claude would need to reproduce the behavior and run the importer to establish whether the change works.

The review should connect those steps:

  • Input: the sample file includes the empty-field case that caused the problem.

  • Change: the correction preserves the intended rows without weakening required-field validation.

  • Evidence: the new test and relevant existing tests run, with failures or skipped checks reported.

A passing test is useful only if it checks the behavior you actually wanted. Inspect the assertion and the code diff, then send a correction if the implementation solves a different problem. You can continue that review in the same conversation.

GitHub pull request Files changed tab with the Diff view menu showing Unified, Split, and Hide whitespace options.

Source: GitHub Docs

For a job like this, the output is a code change supported by execution evidence. The environment has done its job when you can assess that result. The next question is whether your own task can reach that point using the runtimes and tools available there.

Where the Hosted Environment Stops Matching Your Task

Anthropic's standard hosted sessions currently run on Ubuntu 24.04 with an x86_64 architecture. Your laptop's operating system does not change that environment. A Windows installer, a macOS application, and a Linux command-line tool therefore present different compatibility questions.

Start by checking the application that must execute the work, then the resources and access it requires.

Requirement

Anthropic-hosted session

Full cloud computer

Run a compatible Linux tool or test suite

Can run with the necessary dependencies

Can run on a compatible configured system

Run a native Windows desktop application

Outside the supported Ubuntu runtime

Requires a Windows computer and compatible application installation

Perform a GPU-dependent operation

A suitable GPU allocation must be established separately

Choose a machine with the required GPU and driver support

Inspect and edit the application directly

Commands and session interaction go through Claude

Enter the desktop and operate the application yourself

Check the Runtime Before Trying Another Prompt

For example, Adobe lists Windows and macOS requirements for Premiere. A script that changes a Premiere project still needs an appropriate execution path into Premiere. Installing a Python package does not supply the editor, its plugins, or its application APIs.

Adobe Premiere Pro home screen with New Project, Open Project, and a list of recent projects.

Source: Adobe

A useful distinction is where a failure occurs. If an import script produces the wrong values, Claude can investigate its logic. If the required application cannot run on the host's operating system, rewriting the script does not resolve that mismatch. You need to change the execution environment or choose a supported alternative method.

Some applications also offer headless operation. That can be enough for a conversion or scripted calculation. Whether it covers your task depends on which application functions the operation uses and how you will verify the output.

Confirm Capacity for the Actual Workload

Anthropic currently documents approximate resource ceilings of 4 vCPUs, 16 GB RAM, and 30 GB disk. These figures can change. The documentation does not establish a selectable GPU configuration for a rendering workload.

Estimate around the working project: source files, extracted assets, installed tools, temporary data, and outputs all need room. A small script can launch an operation with much larger memory or storage demands. For a GPU job, verify the device, available video memory, driver, and application support before committing a full project.

Separate Saved Setup from an Active Application Session

A cached environment can retain installed dependencies, but its cache does not preserve running processes. Hosted VMs can also pause when idle and later be reclaimed. Plan long operations around the documented session and command behavior.

Direct review introduces another requirement. If you need to select a feature, inspect a timeline, or adjust a viewport yourself, you need access to the application interface. Decide whether returned files and test results are enough, or whether finishing the task requires entering the working application.

A Browser Tab Can Control Different Execution Machines

Claude's web interface can show a task running on Anthropic's infrastructure or connect to a session on another computer. The interface you use to send instructions and the machine that executes them are separate choices.

Remote Control Keeps Execution on Its Host

With Remote Control, Claude Code continues running on the machine where you started it. The browser or mobile app gives you access to that conversation. Commands use the host's files, tools, network, and configured MCP servers.

Terminal running claude remote-control with a connected session and a claude.ai browser link.

Source: DataCamp

If the session starts on your laptop, those operations still use your laptop. Starting Remote Control does not move its application processes to Anthropic's hosted environment.

Claude Code browser session with the connected computer status icon highlighted beside a remote session.

Source: DataCamp

You can instead start Claude Code inside your Vagon computer and connect to that session with an eligible Claude account. In that arrangement, the cloud computer executes the commands. File paths and application bridges must refer to resources available there.

Remote Control and desktop access serve different purposes:

  • Claude's interface: send instructions, follow progress, and respond to the assistant.

  • The Vagon desktop: open the application, inspect its working state, and make changes yourself.

For work to continue, the execution host and the required processes must remain active. Closing an application or stopping that computer affects the task even if you can still view its conversation elsewhere.

Self-Hosted Environments Are Another Execution Route

Anthropic also offers self-hosted environments, currently in public beta for Team and Enterprise organizations. A selected environment routes sessions to runners the organization operates. Its team prepares the tooling, network access, and infrastructure those sessions use.

Consequently, the Ubuntu configuration and resource ceilings described earlier apply to Anthropic-hosted sessions. They do not describe every session visible in the browser. Before troubleshooting a missing application or inaccessible file, identify the selected environment and confirm which machine actually holds the project.

Writing a CAD Script and Working in the Open Model Are Different Tasks

Suppose you want to move a mounting hole in an existing FreeCAD bracket. The instruction sounds simple, but the useful change depends on how the part was built. Which sketch positions the hole? Is its location dimensioned or constrained through other geometry? Which later features depend on it?

FreeCAD sketch with green geometry and red dimensional constraints showing lengths and an arc radius.

Source: FreeCAD documentation

Those questions concern the document's current structure. Generating a new bracket from a description would be a different task and could discard relationships you intended to preserve.

Connect to the Document That Is Actually Open

A community project, theosib's FreeCAD MCP Server, provides a bridge between Claude Code and an addon inside FreeCAD. Its documented tools expose open documents, feature dependencies, object properties, sketch diagnostics, and viewport captures. It also supports executing Python in the application and reporting changes in errors after recomputation.

The arrangement requires both components: an MCP server Claude Code can launch and an addon running in the target FreeCAD instance. We have not tested this integration on Vagon, so verify it with your selected application and Python versions before editing a project.

For the bracket, begin with a request such as:

Inspect the open bracket document without changing it. Identify its file path, the sketch controlling the mounting hole, the constraints that position it, and the features that depend on it. Explain which parameter you would change to move the hole while preserving its diameter.

The expected first result is an explanation grounded in the document. Compare the reported sketch and parameter with FreeCAD's feature tree and property panel. If they do not match, correct the document selection or investigate the model structure before authorizing an edit.

FreeCAD property editor for a selected cube showing placement coordinates and length, width, and height values.

Source: FreeCAD documentation

Application State Changes What the Assistant Can Check

A source file may show the dimensions a script intends to create. Application inspection can reveal whether the current document contains the expected objects and whether its sketches report unresolved constraints. After an approved change, recomputation can expose downstream errors that a plausible script alone would not establish.

Viewport capture adds visual context, while direct desktop access lets you select the feature and examine it yourself. These checks serve different purposes: a clean-looking image does not establish that the model remains editable or dimensionally correct.

FreeCAD also documents headless operation. That remains useful for compatible scripted jobs. Here, the requirement is to work with an existing open document and its application state. The execution environment must provide that document, the running application, and a working connection between them.

Give Claude Code and Your Applications Their Own Vagon Computer

For work that depends on an installed desktop application, we recommend preparing a Vagon computer for Claude Code. You can choose a compatible operating system and hardware, install the application's tools, and keep that setup for later revisions. The FreeCAD example gives you a small first project to prepare.

Start with the Applications the Task Needs

We offer Windows and Ubuntu computers. Choose around the production application and its integration requirements. For a Windows setup, perform the following installations inside the cloud desktop so the application and assistant use the same computer.

Our Computer Composer preinstall catalog lets you request available applications when creating the computer. Supported workflows include Blender, Premiere Pro, and other production software. Selecting those applications saves the separate installation step; accounts, paid licenses, and application connections still need configuration.

For the FreeCAD workflow here, install your selected release from the project's official distribution and launch it before adding the bridge. You can install compatible tools beyond the Composer catalog, so the workflow does not depend on every component being a preinstall option.

Install Claude Code and Verify the Application Connection

Follow Anthropic's setup instructions in a terminal inside the cloud computer, then authenticate with your own Claude access. Prepare the FreeCAD addon and MCP server using the community project's instructions. Configure executable and project paths for this machine.

Linux terminal confirming successful Claude Code installation and showing its version and executable location.

Source: DataCamp

Before loading valuable work, open a sample document and ask Claude to list the open documents. Compare its response with the FreeCAD window. This checks the full route from assistant to bridge to application, rather than only confirming that a server definition exists.

Once that connection works, bring in the bracket project. Keep Claude Code's working folder and the application's document path aligned so both refer to the intended copy.

Transfer the Project, Then Work from the Internal Disk

Upload the model and dependencies through Vagon Files. Dashboard transfers are available while the computer is stopped. Start the computer and sync the project onto its internal disk before opening it.

Use that disk for active editing, reports, and saved revisions. Move completed deliverables back to Vagon Files when you want to transfer or download them.

Your installed applications, settings, and saved files remain between sessions while your subscription stays active. If the workload changes, save and stop the computer before switching performance. The setup stays in place when you restart on the selected tier.

For the CAD inspection, match CPU and memory to the model. For a later GPU-dependent operation, choose suitable GPU resources; we handle GPU driver maintenance. Changing hardware between sessions does not require rebuilding the application environment.

Try One Application Task Before Moving the Whole Workflow

Make the first trial small enough to review in one sitting. For the bracket, use a copy of the document and agree on the mounting hole's new position, its unchanged diameter, and the features that must remain intact. Use the inspection report from the earlier example to confirm what Claude proposes to edit.

After approving the change, judge the result against three checks:

  1. The intended change is present. Measure the hole's position and diameter in FreeCAD. Compare them with the agreed values.

  2. The document still works. Recompute it, inspect any reported errors, and check the dependent features. A successful command alone does not establish a usable model.

  3. The saved revision contains the result. Save under a separate name, close it, and reopen that file. Confirm the geometry and feature tree before preparing delivery.

Keep the original, the revised document, and a short change report together. Record anything that still requires your judgment, so the next session starts with a clear task rather than another investigation.

FreeCAD viewport showing linear measurement components of 17 mm, 9 mm, and 5 mm between selected points.

Source: FreeCAD documentation

Track setup, execution, and review time separately. Our computer usage is billed by the minute, with storage and your Claude access charged separately. Those timings give you a practical basis for estimating a larger job and identifying which preparation can be reused.

Before stopping the computer, save the project and transfer any required deliverables. For unfinished processing, check automatic shutdown settings and leave the necessary processes active.

Create your Vagon computer and begin with one application task whose saved result you can verify.

FAQs

Do I Need to Install Claude Code to Use It on the Web?

You can start an Anthropic-hosted session from claude.ai/code in your browser without installing Claude Code on your device. You need eligible account access and the repository connection required by your chosen workflow. Installing Claude Code becomes relevant when you want its commands to execute on a computer you prepare yourself.

Why Does an MCP Server Work Locally but Fail in a Hosted Session?

The hosted machine may lack the server definition, executable, dependencies, credentials, or connection to the application. Anthropic documents which configuration carries over. In particular, localhost refers to the machine making the connection. A server expecting an application on your laptop cannot find it through the hosted VM's localhost address.

Will a Cloud GPU Improve Claude's Responses?

A GPU on the execution computer can accelerate supported application operations, such as rendering. Claude's model responses come from the model service; adding workstation GPU capacity does not directly speed up that remote inference. Choose hardware for the software and processing the task actually runs.

Can I Keep Using Hosted Sessions After Adding a Cloud Computer?

Yes. Use hosted sessions for suitable repository tasks and the prepared computer for application work. When handing code between them, identify the branch or commit being tested and keep project assets and dependencies aligned. An application's open document, unsaved changes, and running processes need their own handling; they do not move with a Git commit.

Vagon Computer

Run it on a cloud computer

Open it from any browser, Mac, iPad or Chromebook. No local hardware needed.

Vagon Computer

Run it on a cloud computer

Open it from any browser, Mac, iPad or Chromebook. No local hardware needed.

Written by

Mert Boz

Mert is a philosophy student with a strong interest in economics, motorsports, and basketball. He enjoys thinking about how the world works, both in theory and in real life. In his free time, you can often find him watching a race, following a basketball game, or reading about economic ideas.

More posts by this writer →

Related

More on AI Agents

All posts →

AI Agents

Comparison

Choose Windows 365 for Agents or a GPU Desktop

·

13 min read

AI Agents

Fix

Fix MCP Server Not Connecting in Claude Code and Codex

·

13 min read

AI Agents

Comparison

AI Sandbox vs Cloud Desktop: Which Your Agent Needs

·

13 min read

Vagon Computer

Run this guide on a cloud computer.

Open it from any browser, Mac, iPad or Chromebook. No local hardware needed.

Vagon Computer

Run this guide on a cloud computer.

Open it from any browser, Mac, iPad or Chromebook. No local hardware needed.