In action
Watch an agent use its GPU workstation
One request in plain words. Your agent plans the run, rents the GPU it needs, does the work and reports back, all in the same chat.
Agent GPU Sandbox
AI agent sandbox: your agent's GPU workstation
Give Claude, Cursor, Codex or any MCP client its own GPU VM. Your agent deploys an H100 or B200, runs commands, trains and benchmarks models, reads and writes files, and reports back in your chat. You keep SSH access the whole time.
- GPUs
- H100 to B300
- Per command
- 24 h
- Agent client
- Any MCP
Ask your agent
Fine-tune Qwen3 8B on my support tickets and tell me the eval loss.
Spheron connected
Connect once
claude mcp add --transport http spheron https://connect.spheron.ai/mcpRun it in your terminal, then sign in once in the browser.
02 / Deploy
It picks a GPU and asks before it spends.
Your agent searches live offers, chooses a GPU that fits the job and shows you a deploy preview with the hourly price. Nothing is created until you say yes.
03 / Work
It works on the machine like you would.
Shell commands with sudo, package installs, training runs in the background with the log streaming back, files read and written. The VM is yours, so models and data stay between jobs.
04 / Report
It reports back and hands the GPU back.
Results arrive in your chat with the numbers that matter, and your agent offers to Stop or terminate the instance so you only pay for the hours the work took.
Capabilities
What your agent can do on its GPU sandbox
A full GPU VM, driven from your chat. Everything you would do over SSH, your agent can do for you.
- run_commandRun any command
- Shell commands as the login user with sudo. The exit code and full output come back to your agent.
- background: trueKeep long jobs going
- Training and benchmarks run in the background for up to 24 hours, with output streaming back.
- read_file, write_fileRead and write files
- Configs, scripts and results move between chat and machine. Big datasets download on the VM itself.
- kill_commandStop what is running
- Your agent can kill a stuck command and everything it started, then rerun it with a fix.
- VM diskKeep its state
- Packages, downloaded models and checkpoints stay on the VM between tasks, like your own workstation.
- Your SSH keyShare the machine with you
- Your own SSH key works the whole time, so you can log in and see exactly what your agent is doing.
Use cases
GPU work you can hand to your agent
Jobs that take hours on real hardware, not snippets that finish in a second. Ask in plain words.
Fine-tuning and LoRA
“Fine-tune Qwen3 8B on my tickets and report the eval loss.”
Ask for a LoRA or full fine-tune on your dataset and get the eval numbers back, with no script to write.
Model and serving benchmarks
“Benchmark vLLM on two H100s and tell me the tokens per second.”
Benchmark vLLM or SGLang on one GPU or several and get tokens per second and latency in your chat.
Evals on open models
“Run MMLU and GSM8K on these three checkpoints and compare them.”
Run an eval suite across checkpoints or models and get the scores side by side in one table.
GPU test suites
“Clone my repo and run the CUDA tests that fail on my laptop.”
Clone a repo, install its requirements and run the CUDA or PyTorch tests a laptop cannot run.
Data prep and embeddings
“Embed these five million rows and save them as Parquet.”
Clean a dataset, tokenize it or embed millions of rows on a GPU, and keep the output on the VM.
Debugging CUDA and kernels
“This kernel crashes on Hopper. Find out why and fix it.”
Reproduce a GPU only bug, read the profiler output and test the fix on the same hardware.
Compared
A GPU workstation, not a throwaway code sandbox
Code sandboxes are built for quick snippets. The Agent GPU Sandbox is built for the long GPU work an agent cannot do anywhere else.
| Feature | Typical code sandbox | Agent GPU Sandbox |
|---|---|---|
| Hardware | Small CPU containers | H100, H200, B200, B300 and more |
| How long it lives | Seconds to minutes | As long as you keep the VM |
| Start time | Under a second | A few minutes, like any VM |
| State between tasks | Thrown away | Packages, models and files stay |
| Your access | None | SSH with your own key |
| Best for | Running a short snippet | Training, fine-tuning, benchmarks and long GPU work |
How it works
From one question to a GPU at work
- 01
Connect Spheron
Add the connector to Claude, Cursor or Codex and pick a spend limit.
- 02
Ask for the work
Describe the job in plain words. Your agent plans the GPU run.
- 03
Approve the preview
Check the GPU, SSH key and hourly price, then say yes to deploy.
- 04
Get the results
Your agent reports back and offers to Stop or terminate the VM.
Safeguards
Your GPU, your rules
Your agent gets a machine to work on, never a blank cheque.
A preview before anything costs money
Every deploy, Stop, Start and terminate shows a preview first, and your agent waits for your yes.
A spend limit per connection
Cap what a connection can run per hour. A launch over the limit is refused, and only you can raise it.
Only you
Only the person who deployed the VM can use its sandbox. Team members and other accounts cannot.
No open port
The program on the VM only connects outwards to Spheron over HTTPS. Nothing listens on the internet.
Every command recorded
Each command is logged with its text and exit code. Its output and your files stay out of the log.
Remove it any time
Ask your agent to remove the sandbox. Your VM, your SSH access and your running jobs stay as they are.
FAQ / 09
Questions about the Agent GPU Sandbox
What it is, which agents and GPUs it works with, and what it costs.
An isolated machine where an AI agent can run code, install packages and work with files without touching your laptop or production systems. On Spheron the sandbox is a full GPU VM that belongs to you, so your agent can also train, fine-tune and benchmark models on it.
Most code sandboxes are small CPU containers that start in under a second and are thrown away after a snippet runs. The Agent GPU Sandbox is a real GPU VM your agent keeps for hours: it takes a few minutes to boot, and in return it has H100 to B300 GPUs, keeps its state and gives you SSH access.
Claude, Cursor, Codex CLI, VS Code and any MCP client that supports remote servers. Connect Spheron once from Settings, AI connections, and the agent can deploy a GPU sandbox when it needs one.
The same GPUs you can rent on Spheron, from RTX cards and A100s to H100, H200, B200 and B300, single GPU or multi-GPU, on most offers. Your agent shows the GPU and price in the deploy preview.
You pay the GPU instance's normal hourly price until you Stop or terminate it, exactly as without the sandbox. Your agent shows the price in the deploy preview before anything is created.
Yes. Your own SSH key is added like on any Spheron instance and keeps working before, during and after your agent works on it.
A single command can run for up to 24 hours in the background while its output streams back. The VM itself runs until you Stop or terminate it.
Your agent can act only on the VM it deployed, only while you are its owner, and every spend goes through a preview and your connection's spend limit. The program on the VM opens no port, and every command is recorded.
Ask your agent to remove the sandbox. The program is uninstalled and its access revoked, while the VM and your SSH access stay. Terminating the instance removes everything.
// Agent GPU Sandbox
Give your agent its own GPU
Connect Spheron to Claude, Cursor or Codex, ask for the work, and approve the GPU. Your agent does the rest.
