FootprintAI/Containarium

curl -fsSL https://containarium.dev/install.sh | sudo bash   # on a fresh Ubuntu VM, installs Containarium + Incus
sudo containarium create alice --ssh-key ~/.ssh/id_ed25519.pub   # create a box
containarium ssh-config sync   # then add one line to ~/.ssh/config: Include ~/.containarium/ssh_config

Usage

• •Point your agent at the box by adding an MCP entry: command ssh, args ["alice", "agent-box"] in ~/.cursor/mcp.json or ~/.claude.json
• •Agents call shell_exec, read_file, write_file, list_directory, move_file and delete_file inside the container
• •Expose a container port publicly: containarium expose-port alice --container-port 8080 --domain blog.example.com
• •One-command setup: containarium quickstart alice --server <your-vm> --prompt 'a coffee-shop landing page'

Key features

• ✓Persistent, isolated, real Linux boxes (systemd, real networking, open-internet hosting) with blast-radius-bounded credentials (SSH key, not a kube-apiserver token)
• ✓Two MCP surfaces: agent-box (in-the-box shell_exec/file operations with 256 KiB output cap) and mcp-server (create/list/delete/start/stop containers, expose_port, metrics)
• ✓Sentinel architecture (sshpiper + Caddy + PROXY-protocol) survives spot-VM termination and holds the static IP/DNS
• ✓CI-context and CI-prompt resources for agentic debugging of failed CI runs via the containarium-run GitHub Action
• ✓CLI with the same surface plus routes, passthrough, JWTs and system info; Apache-2.0

Best for

AI-engineering teams that want a self-hostable, isolated runtime where agents can build, install, deploy and verify on real infrastructure without touching the human laptop or cluster.