Sandboxes for AI agents,
on your Kubernetes cluster.

Agents get full control inside their sandbox and reach only what you allow outside it. prokube Agent Sandboxes, preconfigured and operated by us.

Use cases

Coding agents for your developers

Claude Code, Codex, opencode or any other coding agent can run with full permissions in a sandbox, without reaching your databases or API keys. Developers connect from the browser or from their own tools.

A place for the agents you build

Your agents get programmatic access to short-lived sandboxes through the Python SDK or MCP. Creating, suspending and resuming one usually takes under a second, and suspended sandboxes use no CPU or memory.

Why sandboxes

Why not on a laptop or a cloud service?

On a laptop, an agent has the developer's access and keys, even inside a container. With a hosted sandbox service, your code and data end up on someone else's servers. A sandbox on your own cluster avoids both.

Agent incidents and sources from our talk

What you get

Customized for you. Run by us.

Sandboxes are part of prokube AgentOps. We install them on your infrastructure and work out the network rules with your security team. After that, we run them.

Where your data is

On-premises, in a private cloud, with a sovereign hosting provider or air-gapped. The snapshots are stored on your cluster too.

Rules you set

Your admins decide which hosts sandboxes can reach and which API keys they get. They can see every sandbox in every workspace.

Operated by prokube

We install updates and monitor capacity. Support with agreed response times is part of the contract.

Running in days

Setup usually takes less than a day. Then we hold a workshop with your team and build the first use case together.

For developers

Control sandboxes programmatically

There is a Python SDK, a CLI called pk and an MCP server. In our tests, a new sandbox answered its first command after about half a second.

from prokube.sandbox import SandboxClient

with SandboxClient() as client:
    client.create("analysis")
    client.write_files("analysis", {
        "/workspace/sales.csv": sales_csv,
        "/workspace/report.py": report_py,  # written by your agent
    })

    client.run("analysis", "python3 -m venv .venv && "
                           ".venv/bin/pip install pandas matplotlib")
    result = client.run("analysis", ".venv/bin/python report.py")
    print(result.stdout)

    chart = client.read_file("analysis", "/workspace/chart.png")
    client.delete("analysis")

Run code in a sandbox

Create a sandbox from a preset and run commands in it. Commands run as root, so the agent can install whatever it needs. Files go in with write_files and come back with read_file.

The UI