One command teleport
for agents, in parallel.
Move your project into a dedicated VM — local or in the cloud — and launch Claude Code, Codex, or Open Code inside it. Each box is isolated, checkpointed, and runs on hardware you control.
Orchestrate agents in loops
liveThe best builders don't babysit one agent — they run many in a loop.
With the AgentBox skill, your agent spawns subagents in isolated VMs and drives them: waits on their input, reads back plan questions, answers, hands over prompts and commands, and polls status. Full git control per box: switch branch, pull, push, open a PR directly from the host. Also with Notion and Linear integration.
Your whole setup, baked in
docker · sandboxStanding up a remote dev box — and keeping it alive — is the tedious part.
AgentBox does it for you. It carries your shell & editor settings, your Claude Code and Codex auth, and your skills and plugins, then provisions the machine end-to-end: Linux packages installed, Docker services up and running, your project bootstrapped and ready. Agentbox.yaml for repetible runs.
Like it never left your laptop
full machineConnecting to a remote box usually means a stripped-down SSH session.
AgentBox gives you the whole machine: a live browser you can screen-share into, a local web tunnel so previews open at your-app.localhost, and a full VS Code / Cursor running inside the box that your local editor attaches to — indistinguishable from local. Plus direct integration with your IDEs and terminals like cmux or Herdr.
How it works
Move your project to a dedicated VM — local or in the cloud — with a single command.
Brings all your skills, plugins, and settings for Claude Code, Codex, and Open Code.
Dedicated browser, screen sharing, persistent shells and a warmed-up VS Code / Cursor IDE in every box.
Sub-1s startup from a previous checkpoint. Auto-pause to save cost and resources when idle.
Git credentials stay on your local machine. Pushing to a remote always goes through an explicit permission request.
The CLI
agentbox createCreate and start a new agent box.agentbox claudeCreate a sandboxed box and launch Claude Code in a detachable tmux session.agentbox <provider> claudeLaunch sandboxed boxes on Vercel, Hetzner, Daytona or local Docker.agentbox <provider> codexSame, for Codex and Open Code — any agent, any provider.agentbox attach 2Reattach to the running agent session on box 2 (claude, codex, or opencode — whichever is live).agentbox urlOpen a box's web app URL in the browser.agentbox screenOpen a box's VNC (noVNC) viewer.agentbox codeOpen a box in VS Code or Cursor via Dev Containers.agentbox shellOpen an interactive bash shell in a box.agentbox dashboardBox list plus the selected box's live agent session.start / stopStart or stop a box, preserving its upper volume.pause / unpauseFreeze or resume a box — sub-second.destroyDestroy a box and discard its upper volume.checkpointList and manage warm box-state checkpoints.agentbox queueQueue new agents to start automatically as running boxes go idle.agentbox agent wait-forBlock until an agent is idle, input-needed or plan-ready.agentbox agent get-planRead an interactive session's plan questions as data — no screen scraping.agentbox drive promptDrive y/N answers, or type straight into any CLI agent's prompt.Providers
local or remote| capability | local docker | hetzner | daytona | vercel |
|---|---|---|---|---|
| support | full | full | partial | full |
| base image | Dockerfile | setup script (Ubuntu) | Dockerfile | setup script |
| live snapshots | yes | yes | experimental | yes |
| private preview urls | portless · orbstack | portless | native | native |
Security architecture
interactiveInstall
The first agentbox create builds the agentbox/box:dev image (~1 GB, one-time). Uses portless to give box web apps the same URL inside the box and on the host.
requirements
- › macOS or Linux — arm64 or Intel
- › Docker — Desktop or OrbStack
- › Node >=20.10
