# Containarium [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/FootprintAI/Containarium  
**GitHub Stars:** 310  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/containarium

## Description
Open-source agent runtime — SSH-native isolation, eBPF egress policy, Kubernetes + LXC backends, GPU passthrough, MCP-native CLI

## Claude Desktop Quick Installation
No confirmed install details yet, so no config is published. Follow the setup steps in the repository above.

## Documentation

## What Containarium MCP server does

The Containarium MCP server provides an administration interface for persistent Linux boxes used by AI agents. It can create and list boxes, configure SSH access, install the in-box agent runtime, and expose a service running inside a box through a public hostname. Boxes are intended to be separate from the operator's laptop and from the host or cluster running Containarium.

Containarium uses unprivileged containers and supports LXC and Kubernetes backends. The runtime can also support GPU passthrough and an opt-in, per-tenant eBPF egress policy. Egress denials are audit logged. A model gateway is available to keep provider API keys outside the box.

## How it works

The platform exposes its administrative surface over MCP. The `containarium` CLI communicates with the local daemon or an HTTP endpoint using a server address and admin token. Commands such as `create`, `list`, `ssh-config sync`, `code install`, and `expose-port` manage the box lifecycle and connectivity.

Each box can be reached over SSH and retains its state between agent runs. The `agent-box` MCP server is installed inside the box, where the agent can use `shell_exec` and edit files directly. An agent therefore receives structured MCP access to the box rather than a Kubernetes API credential. The documented quickstart can also write SSH configuration and connect the box to an agent configuration.

## Setup and configuration

The documented self-hosted path starts with a fresh Ubuntu 24.04 VM where the operator can use `sudo`. The installer installs and starts the daemon, places the administrative token at `/etc/containarium/admin.token`, and adds the CLI to the path. The CLI can then be pointed at the daemon with `CONTAINARIUM_HTTP=true`, `CONTAINARIUM_SERVER`, and `CONTAINARIUM_TOKEN`.

Running `containarium quickstart <name>` creates a box, prepares managed or existing SSH keys, writes `~/.containarium/ssh_config`, and can configure Claude Code by default. `containarium code install <name>` installs `agent-box` inside the selected box. Laptop-based setups can use an SSH tunnel to the VM and synchronize configuration with a jump host. Cursor requires a manual MCP entry that starts `agent-box` through SSH; the documented command also sets the remote executable path.

## Tools and capabilities

- Create and list persistent agent boxes.
- Synchronize SSH configuration and connect through a VM jump host when needed.
- Install the in-box `agent-box` MCP server.
- Run commands with `shell_exec` inside a box.
- Edit files from the agent environment.
- Expose a box port through a public hostname.
- Use LXC or Kubernetes as the runtime backend.
- Apply optional tenant-level eBPF egress policy and audit denied traffic.

## Limitations and notes

Containarium MCP server is self-hosted on the documented path and requires an Ubuntu 24.04 VM with `sudo`; the excerpt does not define installation requirements for other operating systems. The agent-box server is distinct from the administrative MCP surface and must be installed inside each box. Cursor configuration is manual, while the quickstart specifically wires Claude Code by default. The README mentions OpenCode, Gemini, and Codex as agent options, but does not provide enough client-specific configuration here to describe their full setup.

_Full upstream README: https://allmcps.com/mcp/containarium/readme_