OpenSandbox provides a unified API for running AI agents in isolated environments on local Docker or Kubernetes clusters, supporting command execution, file management, browser automation, and desktop access. The platform achieves high-density sandbox creation using Firecracker microVMs with a P50 latency of 64 ms (serial) and P99 of 125.4 ms under 10 concurrent requests, while preserving memory and disk state through pause/resume functionality. Network isolation is enforced via a unified ingress gateway and per-sandbox egress policies, and a Credential Vault injects secrets into workloads without exposing them to agent code. Resource pools and batch creation target evaluation and reinforcement learning workloads. SDKs are available for Python, Java/Kotlin, TypeScript/JavaScript, C#/.NET, and Go, alongside a CLI (osb) and an MCP server for integration with clients like Claude Code and Cursor. Example integrations cover coding agents (Claude Code, Codex CLI), code interpretation, browser automation (Playwright, Chrome), and desktop environments. Images are published to Docker Hub, GHCR, and Alibaba Cloud Registry; legacy per-component releases include Cosign signatures and provenance attestations, though current umbrella releases have a verified signing gap. The project operates under the Agentic AI Foundation with Apache 2.0 licensing.