Questions about agent sandboxes

What is an AI agent sandbox?

An isolated computer where an AI agent can run code, install packages and use tools without touching your own machine. froog's sandboxes are cloud VMs: each agent gets a full Linux machine with its own kernel, root, disk and network, and you drive it from the froog CLI, the HTTP API or the Python, TypeScript and Go SDKs.

How is froog different from E2B, Daytona, Modal or Vercel Sandbox?

E2B, Modal and Vercel Sandbox are built mostly for running code in sessions with a time limit: E2B runs a sandbox for up to 1 hour (24 on Pro), Modal and Vercel Sandbox for up to 24 hours. Daytona sandboxes can run indefinitely but max out at 4 vCPU, 8 GB and 10 GB of disk. A froog sandbox is a whole VM, up to 8 vCPU, 16 GB and 111 GB of disk, that stays on until you stop it (or until its auto-stop timer runs out: an hour by default, up to 30 days, or never), keeps its disk, and comes with SSH, Docker, Chrome, a desktop and /dev/kvm. It also costs less: a 4 vCPU / 8 GB machine running all month is $20 at froog and about $242 at E2B or Daytona list prices (comparison; side by side: E2B, Daytona, Modal).

Is froog an alternative to Fly.io Sprites or exe.dev?

Yes, it's the same idea: a persistent Linux VM for each agent, reached over SSH. froog adds named snapshots you can fork many copies from, an HTTPS URL for any port, a browser desktop and seven coding agents installed. A 4 vCPU / 8 GB VM all month is $20 at froog, about $240 on Sprites at full use, and $35 for an exe.dev resource pool of the same size (froog vs Sprites).

Can I run Claude Code or Codex in a cloud sandbox?

Yes. Claude Code, Codex, Gemini CLI, OpenCode, Amp, Cursor and Pi are installed in every sandbox. Run one with froog prompt NAME "fix the failing tests" --agent claude, or SSH in and use it yourself. Set your API keys once with froog env set and every new sandbox has them.

Can I fork and snapshot a sandbox?

Yes. A snapshot is taken every hour while a sandbox runs, and stopping one snapshots its disk, which is kept for 14 days. Save a snapshot under a name to use it as a template, fork many copies of it to try several approaches in parallel, or restore an older one in place. Snapshots hold the disk. A start within 10 minutes of a stop resumes the paused VM in about a second with every program still running; after that, programs start fresh.

Can agents run Docker, browsers or nested VMs?

Yes. Every sandbox has passwordless sudo, Docker, Chrome (Chromium on Omarchy) and /dev/kvm, so agents can run containers, drive a browser, start an Android emulator or run their own VMs.

Can I use froog as the code-execution tool for my own agent?

Yes. Give your agent tools that run commands and read and write files in a froog sandbox, one sandbox per conversation or user, through the HTTP API or the Python, TypeScript and Go SDKs. There are complete guides for the OpenAI Agents SDK, Claude Agent SDK, Vercel AI SDK, LangChain and LangGraph, and Mastra.

How much does it cost?

Running time is billed per second: $0.0137/h small (2 vCPU, 4 GB), $0.0274/h default (4 vCPU, 8 GB), $0.0685/h large (8 vCPU, 16 GB). Plans are $10, $20 or $50 a month and come back as usage, stopped sandboxes cost no running time, and the first 7 days are free. See pricing.

How long can a sandbox run?

Until you stop it, or until its auto-stop timer runs out: by default an hour after each start, up to 30 days, or never. froog extend NAME --hours 12 moves the deadline and froog extend NAME --no-auto-stop turns it off. A running agent prompt pushes the deadline back so a stop never cuts one off. During the free trial, auto-stop is required and at most 2 hours.

Does a sandbox have its own IP address?

Each sandbox has its own public IPv6 address. Inbound connections to it are refused except SSH and the ports you open with froog port open NAME PORT (TCP or UDP); making a URL public also opens its port. Outbound IPv4 traffic leaves from the address of the machine it runs on, shared with other sandboxes. For HTTPS with a hostname that works for everyone, use froog host NAME PORT. See Network.

How are sandboxes isolated from each other?

Each sandbox is its own KVM virtual machine with its own kernel, memory and disk. Its network port can't reach other sandboxes, the machine it runs on or private ranges, and it can only send from its own addresses. The VM process itself runs unprivileged and sees only its own files. See Security and isolation.

Which operating systems can I use?

Ubuntu 24.04 LTS (the default), Ubuntu 26.04 LTS, Debian 13, or Omarchy, an Arch Linux setup with the Hyprland desktop. Pick one with froog new NAME --os ubuntu|ubuntu-26.04|debian|omarchy. Every one comes with Docker, a browser, language runtimes and seven coding agents installed.