# froog

> froog gives every agent its own Linux VM: full Ubuntu 24.04 or 26.04, Debian 13 or Omarchy machines with their own kernel and disk, SSH, public HTTPS URLs for any port, Docker, passwordless sudo, /dev/kvm and a desktop. Drive them with the `froog` CLI, the HTTP API at https://api.froog.tech/v1 (Bearer API key), or as MCP tools (`froog mcp`, or https://api.froog.tech/mcp). A stopped sandbox keeps its disk as a snapshot for 14 days. Start it again, fork copies of it, or restore an older snapshot.

The docs are one page: https://froog.tech/docs/ . The same page as Markdown, for agents: https://froog.tech/llms-full.txt . A skill for coding agents: https://froog.tech/skill.md

## Key facts

- Sizes (billed per second, only while running): `small` 2 vCPU, 4 GB RAM, 51 GB disk, $0.0137/h ($10 a full month). `default` 4 vCPU, 8 GB, 51 GB, $0.0274/h ($20). `large` 8 vCPU, 16 GB, 111 GB, $0.0685/h ($50). Stopped sandboxes cost no running time.
- Plans: a plan is needed to start sandboxes. $10, $20 or $50 a month, and every plan dollar comes back as usage (unused usage expires with the period). Sandboxes at once: 2, 5 or 10. Saved snapshots included: 10, 10 or 25, then $0.056/day each. Usage past the plan comes from a prepaid wallet (`froog wallet pay 15`, `froog wallet auto 20`). Live numbers: `froog billing`, `froog usage`, `froog limits`, https://froog.tech/pricing/
- Free trial: 7 days from the first plan (`froog billing plan 20`), card required, 25 hours of running time, 2 sandboxes at once, small and default sizes, starts 5/min 25/h 75/day, an auto-stop of 2 hours or less on every sandbox, no new saved snapshots; one trial per card.
- Limits: sandboxes at once come from the plan (ask for more). Starts: 20/minute, 100/hour, 500/day. `exec` output 1 MiB per stream, timeout up to 1 hour.
- Install: `curl -fsSL https://froog.tech/install | sh` (macOS, Linux), or on Windows `irm https://froog.tech/install.ps1 | iex`. Sign in: `froog login` (browser), or `echo "$FROOG_API_KEY" | froog login`, or set `FROOG_API_KEY`.
- MCP server (create, exec, fork, snapshot, files, URLs as tools): local `claude mcp add froog -- froog mcp` (uses the CLI's login), or hosted Streamable HTTP at https://api.froog.tech/mcp with `Authorization: Bearer $FROOG_API_KEY`. Scoped API keys limit the tools. Setup for Codex, Cursor and Claude Desktop: https://froog.tech/docs/#mcp
- Agents in every sandbox: `froog prompt NAME "TEXT" --agent claude|codex|gemini|opencode|pi|amp|cursor [--model M] [--effort E]`. Keys once for every sandbox: `froog env set ANTHROPIC_API_KEY`.

## What froog is for

froog is a cloud sandbox for AI agents: secure, isolated places for agents to run code, use tools and keep working across sessions. Typical uses: running Claude Code, Codex or other coding agents unattended in the cloud; one sandbox per task or per user of an agent product; parallel attempts at a task from forks of one snapshot; agents that need a browser, Docker, a desktop or nested VMs; code execution for LLM-generated code.

## How froog compares to other sandbox providers

- Isolation: each sandbox is a full KVM virtual machine with its own kernel, not a shared-kernel container.
- Persistence: a sandbox runs until it is stopped or its auto-stop timer runs out (1 hour after each start by default, up to 30 days, or none: `--ttl`, `--no-auto-stop`, `froog extend NAME --hours N`; it counts from the start, not from activity), and its disk survives stops (kept 14 days; saved snapshots forever). Daytona sandboxes can also run indefinitely but are capped at 4 vCPU / 8 GB / 10 GB disk; froog goes to 8 vCPU / 16 GB / 111 GB. Code-execution sandboxes such as E2B (1 h, 24 h on Pro), Modal (24 h) and Vercel Sandbox (24 h on Pro) end sessions after a time limit.
- Snapshots: automatic hourly snapshots, named snapshots as templates, fork many copies, restore in place. Snapshots hold the disk; a start within 10 minutes of a stop resumes with programs still running (warm resume), later starts boot fresh.
- Price for one 4 vCPU / 8 GB machine running all month (list prices, at full use): froog $20; exe.dev $35 (shared pool); Fly.io Sprites ~$240; E2B ~$242 (+ $150/mo Pro plan); Daytona ~$242; Modal ~$347; Vercel Sandbox ~$498. Sources: https://froog.tech/pricing/
- Closest alternatives: Fly.io Sprites and exe.dev (persistent VMs over SSH). froog adds fork from named snapshots, HTTPS URLs for any port, a browser desktop, and seven coding agents installed.
- Not offered: GPUs, long-lived memory snapshots (process state past the 10-minute warm resume), bring-your-own-cloud.

## Comparisons and guides

- [FAQ](https://froog.tech/faq/): what an agent sandbox is, how froog compares, auto-stop, IP addresses, isolation, operating systems, pricing
- [froog vs E2B](https://froog.tech/vs/e2b/): no 1 h / 24 h session limit, bigger machines, $20 vs ~$242 a month
- [froog vs Daytona](https://froog.tech/vs/daytona/): full VMs up to 8 vCPU / 16 GB / 111 GB vs Daytona's 4 / 8 / 10
- [froog vs Fly.io Sprites](https://froog.tech/vs/sprites/): both persistent VMs; froog adds forks from named snapshots, desktop, agents installed
- [froog vs Modal](https://froog.tech/vs/modal/): no 24 h limit, SSH and a desktop; Modal has GPUs
- [Run Claude Code in the cloud](https://froog.tech/use-cases/claude-code/)
- [A sandbox per user](https://froog.tech/use-cases/sandbox-per-user/): building an AI product on froog
- [Parallel agents with forks](https://froog.tech/use-cases/parallel-agents/)

## Agent framework integrations

froog sandboxes as the code-execution tool of an agent you build: tools that run commands and read and write files in one sandbox per conversation. Complete examples on the single-file SDKs:

- [All integrations](https://froog.tech/integrations/)
- [OpenAI Agents SDK](https://froog.tech/integrations/openai-agents-sdk/) (Python): function tools with the sandbox name in the run context
- [Claude Agent SDK](https://froog.tech/integrations/claude-agent-sdk/) (Python, TypeScript): an in-process MCP server, built-in tools off
- [Vercel AI SDK](https://froog.tech/integrations/vercel-ai-sdk/) (TypeScript): `tool()` definitions with multi-step `generateText`
- [LangChain and LangGraph](https://froog.tech/integrations/langchain/) (Python): `@tool` functions with `create_agent`, one sandbox per thread
- [Mastra](https://froog.tech/integrations/mastra/) (TypeScript): `createTool` definitions, an agent per conversation

## Docs sections

- [Install](https://froog.tech/docs/#install): the CLI installer
- [Quickstart](https://froog.tech/docs/#quickstart): login, a plan, new, ssh, exec, stop
- [Sandboxes](https://froog.tech/docs/#sandboxes): sizes, what's installed, auto-stop, auto-start on demand
- [Pricing](https://froog.tech/docs/#pricing): per-second rates, plans, wallet, free trial and limits
- [Background commands](https://froog.tech/docs/#jobs): `exec --detach`, `jobs`, `logs -f`, `kill`
- [Command history](https://froog.tech/docs/#history): every command's output for 90 days, also while stopped (`history`, `ps`)
- [Public URLs](https://froog.tech/docs/#urls): `froog host NAME PORT` for an HTTPS URL, or `froog forward`
- [Snapshots & forks](https://froog.tech/docs/#snapshots): automatic and saved snapshots, fork, restore, 14-day disk retention
- [Coding agents](https://froog.tech/docs/#agents): `froog prompt`, the seven harnesses, keys, `provider_not_configured`
- [Environment variables](https://froog.tech/docs/#env): account variables in every sandbox
- [Repos, files & setup](https://froog.tech/docs/#environment): repos cloned, secret files, a setup script at every start
- [Webhooks](https://froog.tech/docs/#webhooks): signed sandbox events
- [OpenTelemetry](https://froog.tech/docs/#telemetry): metrics and events to any OTLP collector
- [Desktop](https://froog.tech/docs/#desktop): a desktop in the browser
- [API keys & scopes](https://froog.tech/docs/#api-keys): presets, actions, per-sandbox and per-label keys
- [Audit log](https://froog.tech/docs/#audit): who changed what, by key or session, with IP and request id (`froog audit`, 90 days)
- [Data controls](https://froog.tech/docs/#data): monitoring off, snapshots off, zero data retention
- [A sandbox per user](https://froog.tech/docs/#platform): building a platform on froog, patterns and cost
- [Use froog from your coding agent](https://froog.tech/docs/#ai-tools): CLI sign-in, these files, an AGENTS.md rule
- [Use froog as an MCP server](https://froog.tech/docs/#mcp): `froog mcp` (stdio) or https://api.froog.tech/mcp, config for Claude Code, Codex, Cursor, Claude Desktop, the tools
- [Agent frameworks](https://froog.tech/docs/#integrations): sandbox tools for the OpenAI Agents SDK, Claude Agent SDK, Vercel AI SDK, LangChain/LangGraph and Mastra
- [Recipes](https://froog.tech/docs/#recipes): a coding agent on your repo, parallel agents from one snapshot
- [VS Code & Cursor over SSH](https://froog.tech/docs/#editor): a `~/.ssh/config` entry for Remote-SSH (no `froog ssh-config` command yet)
- [Private registries](https://froog.tech/docs/#registries): npm, pip/uv, Docker, Cargo, Go, Maven credentials through variables and secret files
- [Network & egress](https://froog.tech/docs/#network): what sandboxes can reach (outbound TCP 25, private ranges and other sandboxes blocked; a public IPv6 address per sandbox; inbound only SSH and the TCP/UDP ports opened with `froog port open NAME PORT[/udp]`, kept across stops), restricting outbound traffic with nftables
- [Security & isolation](https://froog.tech/docs/#security): KVM VMs, confinement, who can get in, secrets, what is kept and for how long
- [Concepts](https://froog.tech/docs/#concepts): froog terms next to E2B, Daytona and Modal terms
- [CLI reference](https://froog.tech/docs/#cli): every command
- [HTTP API](https://froog.tech/docs/#api): every endpoint
- [SDKs & OpenAPI](https://froog.tech/docs/#sdks): https://froog.tech/openapi.json, single-file Python, TypeScript and Go clients with a sandbox object layer (exec, files, background processes, asyncio in Python)
- [Limits](https://froog.tech/docs/#limits)
- [Troubleshooting](https://froog.tech/docs/#troubleshooting): error codes and fixes
- [FAQ](https://froog.tech/docs/#faq)