What it does

AgentTier gives each AI agent or human developer its own sandbox on a Kubernetes cluster. A sandbox is a pod with persistent storage, network isolation and interactive terminal access, and it is declared and managed through custom resources instead of being set up by hand. Built-in templates cover a general coding sandbox, Claude Code and OpenClaw on AWS Bedrock, a Strands Agents setup, a LangGraph agent and a reinforcement-learning rollout worker. Sandboxes can be stopped without losing their files, resumed later, and cloned from a volume snapshot. For code you do not trust, the README describes optional gVisor kernel-level isolation. You deploy it from source with a deploy script, either to a local kind or minikube cluster or to AWS EKS using Terraform. Afterwards you work with it through a web UI with an in-browser terminal, a Python SDK and CLI, or a REST API, and agents can be configured and invoked with streamed output.

Capabilities

One pod per sandbox with persistent storage and network isolation
Optional gVisor isolation for untrusted AI-generated code
Stop, resume and clone sandboxes while keeping their files
Built-in templates for Claude Code, LangGraph, Strands Agents and more
Web UI with a browser terminal, plus Python SDK, CLI and REST API
Deploys to a local kind or minikube cluster or to AWS EKS

When to use it

You want to run coding agents such as Claude Code in isolated environments on infrastructure you control
Your engineering team needs on-demand development environments
You need to execute untrusted AI-generated code with stronger isolation
You are building multi-agent workflows where sandboxes talk to each other