Run AI agents on your own infrastructure. Give them isolated environments to execute code, use tools, and explore the web.
Start with Docker. Scale on Kubernetes.
from opensandbox import Sandboxsandbox = await Sandbox.create("alpine")try: result = await sandbox.commands.run( "echo 'Hello OpenSandbox!'" )finally: await sandbox.destroy()A workspace for every workload
Give agents isolated environments to run tools, manage files, and execute commands.
Let agents navigate the web, test interfaces, and interact with a full desktop.
Execute model-generated code and work with its results through the Code Interpreter SDK.
Run a separate sandbox for each trial, with resource pools and batch creation for parallel workloads.
Fast, high-density sandboxes on Kubernetes. Restore Firecracker microVMs from pre-warmed capacity, and pause with memory and disk state preserved.
Explore Fast SandboxP50 · Serial creation
P99 · 10 concurrent
Your infrastructure. Your boundaries.
One sandbox API, from your laptop to your cluster. Define what agents can reach and how they use credentials.
Route inbound traffic through a unified ingress gateway. Set outbound access policies for each sandbox.
Let agents call external services while keeping real credentials out of sandbox workloads.
Start locally with Docker. Deploy on Kubernetes through a unified, extensible lifecycle and execution API.
From your first sandbox to your next idea
Start the server, pick your SDK, and run your first command. Your infrastructure takes it from there.
Follow the quick start# Initialize your Docker configuration
uvx opensandbox-server init-config \
~/.sandbox.toml --example docker
# Start the server
uvx opensandbox-serverpip install opensandbox