Skip to content

arcboxlabs/arcbox

ArcBox

Sandboxed execution engine for AI agents, containers, and virtual machines.

Built from scratch in Rust -- from hypervisor to CLI.

License Rust Status


Alpha software. ArcBox is under active development. Expect rough edges, missing features, and breaking changes.

Why ArcBox

Computer Use is the next frontier for AI -- agents that can write files, run code, browse the web, and operate a real machine. But giving an agent a full computer means giving it a full attack surface. Containers share the host kernel; a single exploit and the agent is out.

ArcBox solves this with Firecracker-style microVMs that boot their own Linux kernel in under 200ms. Each sandbox is a real computer -- real filesystem, real network, real process tree -- with VM-level isolation that containers can't provide. And when you just need Docker, ArcBox is a drop-in replacement for Docker Desktop.

Three-Tier Runtime

Tier Isolation Boot Time Overhead Use Case
Container Namespace + chroot Instant ~1 MB Standard Docker workloads
Sandbox microVM (own kernel) <200ms ~10-30 MB Untrusted code, CI/CD, AI agents
Machine Independent VM ~1.5s ~200 MB Full Linux dev environment
Host
├── arcbox daemon (Docker API + gRPC)
│
├── System VM (Container + Sandbox tiers, shared kernel)
│   └── arcbox-agent
│       ├── Container Runtime ── namespace + chroot
│       └── Sandbox Runtime ─── KVM microVM (<200ms boot)
│
├── Machine VM "ubuntu-dev" (independent kernel + rootfs)
└── Machine VM "alpine-test"

Sandbox — Computer Use Runtime

Give an AI agent a real computer it can't break out of.

  • <200ms cold boot -- KVM microVM with minimal device model (virtio-MMIO only, no PCI/ACPI/BIOS)
  • <50ms warm start -- snapshot/restore for instant sandbox cloning
  • VM-level isolation -- each sandbox runs its own kernel; a vulnerability in one cannot escape to others
  • Real computer -- real filesystem, real networking, real process tree -- not a simulated shell
  • Disposable -- spin up, let the agent work, tear down; no state leaks between sessions
  • Docker-compatible -- docker run --runtime=sandbox untrusted-image

Container

Drop-in Docker engine replacement. Point your existing Docker CLI at ArcBox:

arcbox docker enable
docker run -d -p 8080:80 nginx

Machine

Full Linux VMs with persistent storage, SSH access, and their own init system.

arcbox machine create dev --distro ubuntu
arcbox machine ssh dev

Quick Start

# Install
curl -sSL https://install.arcbox.dev | sh

# Start the daemon
arcbox daemon start

# Enable Docker compatibility
arcbox docker enable

# Run a container
docker run -d -p 8080:80 nginx
curl http://localhost:8080

What Works Today

  • Container lifecycle -- run, stop, rm, logs, exec, inspect
  • Image management -- pull from Docker Hub and OCI registries (ARM64)
  • Port forwarding -- -p 8080:80 maps host ports into containers
  • Volume mounts -- bind mounts and named volumes
  • Networking -- internet access, DNS resolution, inter-container DNS
  • Docker Compose -- docker-compose up/down for multi-container stacks
  • Context switching -- arcbox docker enable/disable to toggle with Docker Desktop
  • Machine management -- create/start/stop/rm/ls/inspect/exec/ssh
  • 40+ Docker API endpoints -- Docker Engine API v1.43 compatible

Performance

Custom VirtIO stack, zero-copy networking, purpose-built VirtioFS.

Metric Container Sandbox Machine
Boot Instant <200ms cold / <50ms warm ~1.5s
Memory ~1 MB ~10-30 MB ~200 MB
File I/O (vs native) >90% >85% >90%
ArcBox E2B (Firecracker) Docker Desktop
Sandbox boot <200ms ~150ms N/A
Container boot Instant N/A Instant
Idle memory <150 MB Cloud-only 1-2 GB

Known Limitations

Feature Status
docker build Not yet -- use docker buildx or pre-built images
Sandbox runtime (--runtime=sandbox) Designed, not yet implemented
Machine distro management Designed, not yet implemented
x86/amd64 images (Rosetta) Not yet -- ARM64 only
Linux host macOS first, Linux planned
GUI CLI only -- desktop app planned

Requirements

  • macOS 13 (Ventura) or later
  • Apple Silicon (M1/M2/M3/M4) -- Intel support in progress
  • Docker CLI installed (ArcBox replaces the engine, not the CLI)

Contributing

See CONTRIBUTING.md for build instructions, code standards, and development setup.

License

MIT OR Apache-2.0


About

Run AI agents on real and isolated machines — own kernel, filesystem, and network — with <200ms boot. Local first, OCI compatible, pure Rust.

Topics

Resources

License

Apache-2.0, MIT licenses found

Licenses found

Apache-2.0
LICENSE-APACHE
MIT
LICENSE-MIT

Contributing

Stars

Watchers

Forks

Packages

 
 
 

Contributors