Early, before 1.0. It works and is tested (CI on every change, releases with a changelog), but expect rough edges, and breaking changes between minor versions.
Self-hosted sandboxes for your team and their AI agents: everyone gets their own Docker, every app gets a URL, idle sandboxes sleep.
About 10 seconds from New sandbox to your own Docker and a public URL for your app; a sleeping sandbox is back in under 10. (Measured with the images already pulled: the first sandbox on a new server also downloads them.)
Its own Docker daemon for every sandbox (with a Portainer UI if you want one)
A public URL for every app, and TLS addresses for Postgres, Redis and SSH
Sleeps when idle, wakes on the next request; a delete archives, nothing is lost
Made for agents: tokens limited to one sandbox, an agent guide and an MCP server
Why Purgatory, and how it compares with a laptop, a cloud PaaS, a shared Docker host, Coolify and a VM for everyone · How it works: the components (Docker-in-Docker, frp, Traefik, Sablier), a sandbox's lifecycle and the path of a request
All real runs, in real time; press one to watch it.
With curl through the API, by an agent over the MCP server, and the MCP server by hand (JSON-RPC over HTTP, so even curl can call its tools).
With Docker Desktop, or Docker Engine and the compose plugin on Linux, and ports 80 and 443 free (on Windows, in Git Bash or WSL):
mkdir p7y && cd p7y
curl -fsSLO https://github.com/p7y-dev/p7y/releases/latest/download/docker-compose.yml
curl -fsSL https://github.com/p7y-dev/p7y/releases/latest/download/setup.sh | bash # prints the admin password
docker compose up -dOpen http://p7y.lvh.me (lvh.me and all its subdomains point to 127.0.0.1, so no DNS to set up), sign in as administrator with the password setup.sh printed, and press New sandbox. The first one takes a minute or two while the images are pulled; after that about 10 seconds. Its app opens at its own http://<name>-….lvh.me address.
To remove it: delete the sandboxes in the UI first (each is a compose project of its own, so the next step would leave them running), then docker compose down -v in the p7y directory, and delete the directory.
Related MCP server: anvil
On a Linux host with Docker Engine and the compose plugin, no clone needed (check the host first):
mkdir p7y && cd p7y
curl -fsSLO https://github.com/p7y-dev/p7y/releases/latest/download/docker-compose.yml
curl -fsSL https://github.com/p7y-dev/p7y/releases/latest/download/setup.sh | bash -s -- dev.example.com # .env with your domain and random secrets
docker compose up -dThen sign in at http://p7y.<HOST_DOMAIN> (by default http://p7y.lvh.me).
p7y-dev.github.io/p7y: install, configuration, let your agent deploy, MCP, API, and llms-full.txt for agents.
Security: where sysbox is installed, sandboxes run without a privileged container; without it they fall back to the privileged
dind, which suits a team that trusts each other, not hostile code. On a shared server install sysbox and setALLOWED_RUNTIMES=sysbox(details).
Contributing · Security issues · Apache-2.0
Free to use, also commercially. If you ship Purgatory or a product built on it, keep the NOTICE with it, so your users see it is based on Purgatory; the names "Purgatory" and "p7y" are not licensed for another product's name.
This server cannot be deployed
Remote Linux boxes for coding agents: Docker, a browser, screenshots, logs, human takeover.
Hosting for AI agents: your AI client deploys Docker apps to live HTTPS URLs over MCP.
Real Linux labs your AI agent deploys, routes and runs, with domains, TLS, DBs and an audit log.
Persistent cloud workspaces for AI agents: run commands, edit files, use git and a browser.