agent-sandbox: Run Isolated Sandboxes for AI Agents

Overview

Agent-Sandbox is a Go service that lets agents and developers provision isolated environments for untrusted code and actions without managing Kubernetes resources directly. It builds on a Kubernetes sandbox foundation and exposes sandbox operations through a REST API, an MCP server, and a built-in web UI.

It is intended for teams that need per-agent or per-user isolation and persistent sandbox state across interactions. It is a good fit when Kubernetes is available and you want a self-hosted alternative to hosted sandbox services; it may be unnecessary if you only need local execution or do not operate a Kubernetes cluster.

Key Features

• Provides sandbox lifecycle management through REST APIs and an MCP server.
• Supports E2B protocol and SDK compatibility for existing E2B-based integrations.
• Offers environments for code execution, shell access, browser use, and computer or desktop use.
• Stores templates, blueprints, and sandbox state in Kubernetes objects, without requiring a separate database or message queue.
• Includes a web UI for inspecting and managing sandboxes, templates, and pools.
• Supports warm pools, pause and resume, snapshots, and scale-to-zero behavior.
• Separates deployment configuration in Blueprints from sandbox type configuration in Templates.

Use Cases

• AI application teams can run LLM-generated code in isolated, persistent environments rather than on an application server.
• Agent platform operators can provide separate sandboxes for users or agents while managing them through a shared Kubernetes service.
• Developers migrating integrations built for E2B can evaluate a self-hosted service with E2B-compatible SDKs and APIs.
• Teams building browser or desktop agents can provision environments that include browser use or GUI access.

Getting Started

The project requires Kubernetes 1.28 or later. Apply the installation manifest to a cluster:

kubectl create namespace agent-sandbox
kubectl apply -nagent-sandbox -f install.yaml

See the README and documentation site for setup, authentication, API, and SDK details.

Alternatives

• OpenSandbox: OpenSandbox also provides isolated environments for agent tasks, with local Docker support alongside Kubernetes and a shared sandbox API.

Considerations

• A Kubernetes cluster is required, with version 1.28 or later; this is not a standalone local sandbox runtime.
• Sandboxes execute untrusted code and actions, so deployment configuration and access controls matter. The README describes a default admin login token, which should be treated as a setup caveat and secured appropriately.
• The project has 218 stars and 4 open issues in the supplied repository data. These figures are only a snapshot and do not establish production maturity.
• The repository describes broad workload support, but the README's quick-start example focuses on the E2B Python SDK. Check the documentation for the integration and sandbox type you need.