OpenSandbox is a free, open-source runtime that gives AI agents their own isolated sandboxes to run code, execute commands, manage files and drive browsers or desktops. You can start with Docker on your laptop and move the same setup to Kubernetes when you need scale.
OpenSandbox at a glance
Price | Free (Apache-2.0) |
|---|---|
Runs on | Docker locally, Kubernetes for production |
SDKs | Python, Java/Kotlin, TypeScript/JavaScript, C#/.NET, Go |
Extras | osb CLI and an MCP server for Claude Code, Cursor and similar clients |
Requirements | Docker and Python 3.10+ |
What OpenSandbox does
Agents that write and run code need somewhere safe to do it. OpenSandbox provides that through one sandbox API: built-in endpoints for commands, files and code execution, plus examples for coding agents, browsers and desktops. Inbound traffic goes through an ingress gateway, and every sandbox has its own outbound (egress) policy, so you decide which sites an agent can reach. A credential vault lets agents call outside services without the real keys ever entering the sandbox.
On Kubernetes it uses Firecracker microVMs. The project reports creation times of 64 ms at the median (one at a time) and 125.4 ms at the 99th percentile with 10 created at once, and Firecracker sandboxes can be paused and resumed with their memory and disk intact.
What you can do with it
Let a coding agent run tests and scripts without touching your machine.
Give a browsing agent a disposable browser that can only reach approved sites.
Spin up batches of sandboxes for evaluations or reinforcement-learning runs.
Plug sandboxes into Claude Code or Cursor through the MCP server.
How to get started
The README's quick start: run uvx opensandbox-server init-config ~/.sandbox.toml --example docker, then uvx opensandbox-server, then install an SDK such as uv pip install opensandbox. The CLI and MCP server are separate installs (opensandbox-cli and opensandbox-mcp).
What to know
The speed figures are the project's own measurements.
Production use means running and securing your own Kubernetes cluster.
Our take
Sandboxing is the part of agent building most teams improvise, and that's where incidents come from. OpenSandbox packages the pieces you'd otherwise bolt together (isolation, network rules, secrets handling) under a permissive license. If you're letting agents execute code anywhere near production, it's worth an afternoon.
OpenSandbox alternatives
AWS Emulator: a local stand-in for AWS so agents can test cloud code offline.
ZeroClaw: a self-hosted agent runtime with OS-level sandboxing built in.
Paradev: run several coding agents in parallel, each in its own Git worktree.
Frequently asked questions
Is OpenSandbox free?
Yes, it's open source under Apache-2.0.
Do I need Kubernetes?
No. Docker is enough to start, and the same lifecycle API works when you move to Kubernetes.
As featured in World of AI
Google's new API scores answers instead of writing them (Nov 11, 2026)
Facts checked against the official OpenSandbox repository on November 13, 2026.
Want four tools like this every weeknight? Subscribe to World of AI for free. Browse more in our AI Tools directory.