# sandbox

Rust-based agent OS with built-in security, WASM sandboxing, and multi-agent runtime

OpenFang is an open-source agent operating system built in Rust that provides a secure multi-agent runtime with WASM sandboxing, auditability layers, and multi-channel communication. It goes beyond typical orchestration SDKs by treating agent security and operational isolation as first-class concerns, making it suitable for teams deploying agents in environments where trust boundaries and audit trails matter.

Open-source sandboxes and SDKs for AI agents that control desktops

CUA is an open-source infrastructure platform for building, benchmarking, and deploying AI agents that autonomously control full desktop environments. It provides secure sandboxed VMs across macOS, Linux, Windows, and Android with a unified Python SDK for screenshots, mouse/keyboard control, shell commands, and file I/O. Includes CuaBot CLI for running agents in sandboxes, Cua-Bench for standardized evaluation, and Lume for near-native macOS virtualization on Apple Silicon.

Local microVM sandboxes for AI agent code execution

Microsandbox provides hardware-level isolated sandboxes for AI agents to execute code safely on local machines. Using microVMs with sub-200ms boot times, it offers stronger isolation than Docker containers while remaining lightweight enough for development workstations. OCI-compatible with support for Python and Node.js runtimes. Apache-2.0 licensed with 5,000+ GitHub stars.

Open-source browser infrastructure for AI agents at scale

Steel is an open-source browser API purpose-built for AI agents, providing managed headless browser sessions with anti-bot bypass, proxy rotation, CAPTCHA solving, and session persistence. It handles the infrastructure layer that browser automation agents like Browser Use and Stagehand run on top of. Self-hostable or available as a cloud service. Over 6,000 GitHub stars.

Containerized sandboxes for AI coding agents

Dagger Container Use provides isolated container environments for AI coding agents, enabling multiple agents to work in separate sandboxed branches simultaneously. Built by the Dagger team, it ensures reproducibility and safety for autonomous code execution by giving each agent its own containerized workspace with full toolchain access.

Open-source AI coding agent for large projects and real-world tasks

Terminal-based coding engine that breaks complex tasks into subtasks with support for 2M token context windows, diff review sandboxes, and both autonomous and manual modes. Multi-model pipelines work with Anthropic, OpenAI, and Google. With 15k+ GitHub stars, Plandex is a mature open-source option for developers tackling large, multi-step development tasks from the command line.