Skip to content
consoles

A shared workbench
for you and your agents.

Let your agent render a Blender scene on a cloud GPU. See the image, change the controls, and run it again. Keep the code, files, and results in one workbook you can reopen from any device.

Open ConsolesExplore recipes

Your agent writes the code. You can see and rerun what it makes.

Inside a workspaceInteractive example · simulated output
GPU labNVIDIA A102 cores · 8 GiB
NotebookTextCodeViewAbout
AgentCodexjust now
Python
OutputExit 0
Using NVIDIA A10 · CUDA available

epoch 1/3   train_loss 1.842   eval_loss 1.916
epoch 2/3   train_loss 1.327   eval_loss 1.481
epoch 3/3   train_loss 1.094   eval_loss 1.263

Saved outputs/metrics.csv
metrics.csvDownload example
Try Run to replay the example.No compute used

Start with a recipe.
Make it your own.

A Blender studio. A build runner. A place to explore your data. Give your workspace the tools and starting files for the job.

Built-in starters and ideas
for your own setup.

Blender studioLinux · NVIDIA GPU
Built-in Blender recipe

From a brief to a finished render.

Give your agent a scene to make or a .blend file to work on. Review the render, ask for a different angle, and keep the editable scene alongside it.

Blender studio render of a smoked glass sphere on a metal pedestal, lit by blue and coral rings
render.pngscene.blendscene.glbExample outputs
Tools & runtime
  • Blender
  • Python
  • Cycles · OptiX / CUDA
Try this with your agent

Create a Blender GPU workspace. Make a product scene with smoked glass, a metal base, and colored lighting. Render a frame and save the .blend file and a GLB so I can inspect and refine it.

The built-in recipe includes Blender, checks GPU rendering, and starts with a scene your agent can edit.

What else could you make?

Image generation studio

Start from ComfyUI. Add your models, compare image variations, and keep the workflow that made the one you like.

Document factory

Turn project data into PDFs, slides, or recurring reports. Keep the sources, generation script, and finished files together.

Repository maintenance

Upgrade dependencies, migrate a framework, or work through a failing test suite. Give your agent a place to try the changes.

See what it’s doing.
Make it yours.

Create a workspace, connect your agent, and give it a task. The files it changes and commands it runs stay within reach.

A real filesystem
Each Linux workspace has files and folders for your code, data, and outputs. You can browse and edit the same files your agent uses.
An isolated environment
CPU and GPU code runs in its own cloud sandbox, separate from your computer. Install the packages and command-line tools your task needs.
Work that lasts beyond a session
Inspect saved renders and earlier runs after compute stops. Each new Blender run keeps its own output files, submitted controls, and source snapshot.
You and your agents, working together
Connect your MCP agents to the same workbook across sessions and devices. See who ran each command, inspect its output, or edit a cell and run it yourself.

Choose the compute
your work needs.

Start with a Linux CPU sandbox for scripts and debugging, or choose an NVIDIA GPU for more demanding work.

NVIDIA GPUs

Up to 8 GPUs per workspace

  • T4

    16 GB GPU memory

  • A10

    24 GB GPU memory

  • L40S

    48 GB GPU memory

  • A100

    40 GB GPU memory

  • H100

    80 GB GPU memory

Choose 1–8 GPUs (up to 4 A10s), 0.5–32 physical CPU cores, and 1–256 GiB RAM. Customize your machine before starting.

Linux CPU
Up to 64 vCPUs · 256 GiB RAM

Choose your CPU and memory for Python scripts, parallel builds, data processing, and debugging. Starts at 1 vCPU and 1 GiB RAM.

TypeScript
Isolated JavaScript / TypeScript runtime

Run short scripts and call connected tools without starting a Linux machine.

Your tools.
Ready at startup.

Save the setup with the workspace. Install APT packages, restore npm dependencies, or prepare a Python environment before your agent gets to work.

Setup commands run each time compute starts. Your starter files and saved work give the agent a place to pick up.

Create your workspace
Example startup commands
# System tools
apt-get update
apt-get install -y ffmpeg

# Python dependencies from your files
python3 -m venv /data/.venv
/data/.venv/bin/pip install -r requirements.txt

# npm dependencies from your lockfile
npm ci --prefix web
Linux CPU or GPU workspace · customize for your project

Pause the machine.
Pick up where you left off.

Your notes, commands, and recorded results stay in the workspace after compute stops. Return to review a run or start the next one.

Retained storage keeps the files you need between runs. Configurable CPU and GPU workspaces include a persistent /data directory for larger datasets and checkpoints. Running processes and GPU memory do not carry over.

You can stop a machine yourself or set an idle timeout.

A few practical details.

Do I need a particular AI agent?

Bring an agent that supports MCP, the connection used to give agents access to tools. The Connect agent dialog includes setup instructions for Codex. You can also run commands directly in the workspace.

What can my agent do?

With the permissions you approve, an agent can read and write files, add cells, run commands, inspect logs, and retrieve results. You choose whether it has access to one workspace or all your workspaces.

Can my workspace use my APIs and tools?

Add MCP or OpenAPI services in Connections, approve tools for your workspace, and call them from TypeScript cells.

Can I set up my own recipe?

Yes. Start with a built-in template or a blank Linux workspace, then add your files and setup commands. Setup can install APT, npm, and Python packages each time compute starts. Recipes that need a specific engine, SDK, or operating system also need the matching environment.

Does the browser need to stay open?

No. Jobs run on the compute service, independently of the browser tab. Open the workspace again to see their status and recorded output.

Does the entire machine get saved?

No. Notes, commands, and recorded results are saved separately from the machine. Files in /workspace use size-limited snapshots; configurable CPU and GPU workspaces also provide persistent /data storage for larger files. Use setup commands to reinstall dependencies. Running processes and GPU memory are not restored.

Where can I see compute usage?

In a CPU or GPU workspace, open the usage view to see its recorded compute sessions and cost estimates. The amount depends on the machine and how long it runs.