sandboxing

BoundaryBench is an open-source benchmark for coding agents running under hardened sandbox policies. It compares harnesses like Claude Code, Codex, Terminus 2, and Grok on Terminal-Bench tasks and includes quickstart, policy controls, and result export tools.

Open-source AI agent framework and meta-harness for orchestrating Claude Code, Codex, Cursor, Pi, and custom agents with policies, sandboxing, and cross-device collaboration.

Linux command runner that kills every process started by a command and verifies cleanup before returning. Built for AI coding agents and CI jobs, with enforced cgroup mode, fallback process scanning, time and resource limits, and JSON run records.

Legion is an open-source Elixir runtime for AI agents that run inside your app, write code, and act on your data through sandboxed tools. It supports Lua or Elixir code execution, user-scoped actions, Postgres-backed persistence, and a LiveView dashboard for observability.

Sandboxed macOS VMs with full Xcode for AI coding agents, plus a lightweight Linux container mode. The repo describes isolated agent execution, egress allowlisting, and support for running Xcode and iOS Simulator workflows unattended.

Desktop app for running CLI coding agents like Claude Code and Codex in parallel git worktrees, with built-in diff review, multi-repo support, broadcast, and sandboxing.

Spens is a developer tool for running coding agents in sandboxed Docker containers with network controls, secret handling, and detailed logs of LLM calls, tool use, and file changes.

XSAF is a small, fetch-native TypeScript framework for building secure AI agents. The page shows its core API, agent lifecycle, tool pipeline, MCP support, session handling, and offline testing story.

Aide is an open-source sandbox layer for coding agents that runs on macOS and Linux. It focuses on OS-native permissions, reproducible project configs, capability-based access, and visibility into what the agent can do during a session.

Open-source governance harness for AI agents with policy enforcement, human approvals, isolated execution, budgets, and signed audit evidence. It includes a CLI, REST API, web dashboard, SDKs, and adapters for common agent frameworks.

A web utility for creating end-to-end encrypted share links for single-file HTML generated by agents. It stores ciphertext for seven days and lets browsers decrypt and render the page from the URL fragment in a sandbox.

A self-hosted Python runtime for building agentic AI apps with server-driven UI, audit logging, RBAC, sandboxing, and pluggable model orchestration.

An open-source reference implementation for autonomous vulnerability discovery and remediation with Claude. It includes Claude Code skills for threat modeling, scanning, triage, patching, plus a harness for running a recon → find → verify → report → patch pipeline.

A secure Model Context Protocol server that gives AI assistants controlled terminal access with risk analysis, sandboxing, logging, filesystem protection, and optional Docker and Git features.

Open-source Rust coding agent with LLM-native code understanding, shell safety, and support for multiple LLM providers with automatic failover.

A CLI for running untrusted MCP servers in sandboxed containers with deny-by-default egress, traffic watching, and host approval workflows. It can also compose servers into agents and expose them over OCI registries.

Open-source CLI for spinning up disposable local or remote sandboxes for AI agents. It uses a manifest and Dockerfile to build fresh boxes, then exposes a single shell tool through MCP so an agent can work inside the sandbox without touching the host.

A Go-based CLI for managing Python virtual environments, with MCP tools for AI agents, typed JSON snapshots, ephemeral execution, and file-watcher sync for evolving scripts.

A Claude Code skill that launches a second model in a herdr split pane to adversarially review a diff, then checks each finding before returning a verdict.

Teleport-Env is an ultra-fast OS-level snapshot and rollback sandbox for autonomous coding agents, built with overlayfs and CRIU. It targets destructive agent testing, MCTS search loops, and reinforcement learning workflows that need rapid environment recovery.

Docker blog post about a real AI coding agent failure and how Docker Sandboxes aim to contain destructive execution mistakes.