Skip to main content
Kubox is a platform for the agentic era. Run agents and apps inside your trusted boundary, with control over where they run and proof of what happened. It is three products on one platform. Most people start with one.

Coding Sessions

Describe a task and watch an agent work, in an isolated sandbox. The major coding agents are already installed.

Kubernetes Management Plane

One self-hosted management plane controlling hundreds of clusters across separate AWS accounts.

Open-weight Inference

Choose a model, see what it costs, and send a request. In partnership with SCX.

​
Where to start

Call the inference API

Point an OpenAI or Anthropic SDK at Kubox and send your first request. About five minutes.

Build your first cluster

Create a Kubernetes cluster in your own AWS account with the kubox CLI, then tear it down. About 15 minutes.

Understand the pieces

How the console, management plane, cloud accounts, and clusters fit together.

Browse the commands

Every kubox command, its options, and what it changes.

​
Kubernetes, in your account

The management plane builds clusters in your AWS account, or in your customers’. That suits two shapes of team:
  • SaaS providers — run your software in a customer’s own cloud account when they ask for it.
  • Platform teams — hand clusters to internal teams without becoming the bottleneck.
Either way: one plane, independent clusters, and a named owner for each. Kubox builds the cluster. Deploying your application is still your job. A few properties worth knowing before you evaluate it: Architecture explains how these pieces relate, and cloud accounts covers exactly what Kubox can and cannot do inside your account.

​
Inference, with the cost in front of you

Kubox serves open-weight models through an endpoint compatible with both the OpenAI and the Anthropic APIs. Existing code works by changing two lines — see the API quickstart. Name each key after the app that uses it. Set a budget per key. See what each one has spent, and the calls behind it.
Selected models run on SCX infrastructure in Australia. Hosting differs per model, so check the location shown on the model you pick rather than assuming it applies across the catalogue. SCX’s claims about their platform are theirs — read them at scx.ai.

​
Coding Sessions

Give a coding agent a task and watch it work — every step, and what it costs, as it happens. Each session runs in its own sandbox. Your API keys never reach the agent. The agent shows you each command and waits: approve it, stop being asked, or refuse. Coding Sessions is driven from the console. See kubox.ai/coding-sessions until reference documentation lands here.

​
What else runs on it

Kubox clusters are ordinary Kubernetes clusters. Your existing manifests, operators, and tools work unchanged — Ray Serve, Dask, Spark, Dagster, Qdrant. The use cases work through some of these end to end, including a three-billion-row analysis and a $15 satellite-imagery pipeline.