sandbox

Open-source AI sandbox infrastructure for running code, browsing the web, and isolating agent workflows inside disposable microVMs.

Open-source infrastructure for computer-use agents, with sandboxes, SDKs, benchmarks, and desktop automation tooling for macOS, Linux, Windows, and Android. The repo also includes Cua Driver, CuaBot, Cua-Bench, and Lume for VM management.

A research prototype for a BitTorrent-style mesh where coding agents share idle capacity, delegate tasks, and settle credits across peers.

Mirrors turns production traces into an isolated, runnable copy of an AI agent’s environment so teams can replay workflows, reproduce bugs, and test changes safely before shipping.

Tail Panic is a code-driven battle game where you script AI agents to chase or evade on a 25×25 grid, join bot matches, and review 3D replays.

Disposable, network-restricted Linux VMs for AI coding agents. The project is built for letting coding agents install packages, run tests, use system tools, and work inside an isolated machine instead of the host laptop.

SpadeBox is an open-source sandboxed tools and JavaScript runtime for AI agents, with Rust core plus JavaScript and Python bindings. It provides file, network, and code-execution tools, supports MCP, and includes examples for JavaScript, Rust, and Python.

A small Linux CLI for running coding agents in a sandboxed environment. The repository shows installation options, TOML-based configuration, supported agents, and troubleshooting notes.

Open-source, self-hostable sandbox for AI agents. It lets agents work inside isolated Linux containers through MCP tools, with CLI and host-level admin controls for creating, managing, and exposing boxes.

Open-source infrastructure for running and automating sandboxed coding agents, with local Docker setup, integrations, snapshots, sessions, and automations.

Open-source AI-powered desktop automation agent that gives an LLM a fully isolated Ubuntu workspace with browser-based supervision, MCP access, audit logs, and OS-level control.

A browser-based demo where an AI agent writes x86_64 assembly and eBPF to generate Wolfram cellular automata and Mandelbrot fractals inside a real Linux VM, with framebuffer and kernel trace streams exposed live.

A minimal native macOS sandbox for running Claude Code or Codex with OS-enforced limits on what the agent can touch.

A demo for running the Odysseus self-hosted AI workspace on throwaway islo.dev microVMs in one command, with both a clickable live app path and an agent tool for running code on disposable boxes.

Open-source local compute fabric for AI agents. It runs a single-node K3s stack through Docker Compose, provisions isolated Ubuntu sandboxes, and exposes native MCP support for local agent workflows.

A research-style demo showing LLM agents evolving through a genetic algorithm to play Pokémon in parallel sandboxed VMs. The page explains the setup, scoring loop, generations, results, and links to the repo and run artifacts.

Sandboxed bash for agents with structured command output and full filesystem state capture after each run.

A harness that runs MCP servers in a disposable Linux sandbox and logs their real startup behavior, including file access and outbound network connections.

Code Airlock is a wrapper for running coding agents like Claude Code and Codex inside disposable microVMs, with changes committed to git for host-side review.

A Docker-based sandbox for running AI coding agents with dropped capabilities, no-new-privileges, and no Docker socket access. It wraps the pi terminal coding agent and supports local file edits, persistent auth, and skills mounts.

Autolang is a scripting language and compiler for AI agent code, designed to run untrusted AI-written scripts in a controlled environment with restricted access to registered functions.

A blog post describing ptc_runner’s MCP server and code-mode approach, where agents run short-lived untrusted programs inside a small Lisp REPL instead of Python or JavaScript.