Containarium — Agent Runtime

Open-source agent runtime · SSH-native isolation · eBPF egress policy · Kubernetes + LXC · MCP-native CLI · GPU passthrough

The open-source, self-hostable agent runtime for AI agents. Each agent gets a persistent, SSH-reachable box in an unprivileged container, with an opt-in per-tenant eBPF egress policy, audit-logged denials, and a model gateway that keeps provider API keys out of the box. What is on by default, and what each claim rests on, is in docs/security/CLAIMS.md.

Bring your own agent — Cursor, Claude Code, OpenCode, your own MCP client. We run the box.

agent: "create me a sandbox called 'blog'"           → containarium create
agent: "wire up SSH so I can reach it"               → containarium ssh-config sync
agent: "install Caddy on :8080 inside the box"       → shell_exec (via agent-box MCP)
agent: "expose that on blog.example.com"             → containarium expose-port

curl https://blog.example.com → hello world

🌐 Project site: containarium.dev · 🎬 55s demo: youtu.be/IBDDD_tb8FY · 🚀 Live app: helloworld.demo.containarium.dev

---

Choose your path

Path	You end up with	Go to
Hosted cloud	a box you can ssh into with your agent wired to it, no server to run	Sign up at cloud.containarium.dev
Self-host on a VM	the same, on your own Ubuntu VM (about 5 minutes)	Quick start
See the end result	a live app served from a box, which is what you end up with	helloworld.demo.containarium.dev

---

Why an agent runtime?

AI agents are increasingly the primary user of dev infrastructure. They want to build, install, deploy, and verify — not on the human's laptop (too noisy, too risky, too local) but on a persistent runtime that's:

• Persistent: state survives between agent runs.
• Separate from your machine: a misbehaving install doesn't touch your laptop or your cluster. (Tenant-to-tenant isolation is not on by default; see docs/security/CLAIMS.md.)
• Real: a full Linux environment with systemd, real networking, and the ability to host things on the open internet.
• Driven by structured tools: not by an agent typing commands into a TTY hoping nothing scrolls off-screen, but by MCP — typed, bounded, safe.
• Blast-radius-bounded: the agent holds an SSH key, not a kube-apiserver token. It can't reach the cluster control plane, the host OS, or other tenants' boxes.

That's the runtime Containarium gives you. It runs as a self-hosted platform on LXC or Kubernetes, exposes its admin surface over MCP, and ships a second MCP server that lives inside the box so the agent can shell_exec and edit files directly.

You bring the agent. We run the box.

---

Quick start

One path, three steps, with a check at the end of each. You're done when a box exists, ssh alice gets you a shell in it, and your agent can run commands inside it.

Before you start: a fresh Ubuntu 24.04 VM you can sudo on (see System requirements). You need Containarium v0.96.0 or later; the installer fetches the latest release. sudo is only needed to run the installer and to read the admin token. Every other command runs as your normal user and goes through the daemon's API.

1. Install, then create your first box

On the VM:

curl -fsSL https://containarium.dev/install.sh | sudo bash

✅ Check: the daemon is running and the CLI is on your PATH.

systemctl is-active containarium   # → active
containarium version               # → v0.96.0 or later

Point the CLI at the local daemon, using the admin token the installer saved, and create a box called alice:

export CONTAINARIUM_HTTP=true
export CONTAINARIUM_SERVER=http://localhost:8080
export CONTAINARIUM_TOKEN="$(sudo cat /etc/containarium/admin.token)"

containarium quickstart alice

quickstart creates the box, reuses your ~/.ssh key or generates a managed one (~/.ssh/containarium_ed25519), writes ~/.containarium/ssh_config, adds the single Include ~/.containarium/ssh_config line to ~/.ssh/config, and wires the box into your agent's MCP config (Claude Code by default). It's safe to run again, including to repair an older quickstart's MCP command.

✅ Check: quickstart ends with ✓ quickstart complete, and the box is listed as running.

containarium list   # → a row for alice-container, STATUS running

2. Connect to the box

Choose the variant that matches where you're typing.

Variant A: on the VM. This works on a fresh install. Step 1 already wrote the config. Re-sync whenever your boxes change:

containarium ssh-config sync --identity ~/.ssh/containarium_ed25519
# leave out --identity if quickstart reused your own ~/.ssh key

✅ Check:

ssh alice hostname   # → alice-container

Variant B: from your laptop. On a single VM, ssh-config sync adds a ProxyJump through the VM's per-user jump account when the daemon doesn't advertise ssh_host. If the API is reached directly, the jump host defaults to the API hostname, on SSH port 22. With the API tunnel below, name the actual VM with --jump-host (use <vm-host>:<ssh-port> for a custom SSH port). Deployments with a sentinel can use --sentinel <sentinel-host> instead; a daemon's --ssh-host still takes precedence over the jump route.

# On your laptop: install the client only.
curl -fsSL https://raw.githubusercontent.com/footprintai/containarium/main/hacks/install-cli.sh | sudo bash

# Reach the daemon's API through an SSH tunnel to the VM.
ssh -fN -L 8080:localhost:8080 <you>@<vm-host>
export CONTAINARIUM_HTTP=true
export CONTAINARIUM_SERVER=http://localhost:8080
export CONTAINARIUM_TOKEN="$(ssh <you>@<vm-host> sudo cat /etc/containarium/admin.token)"

containarium ssh-config sync --jump-host <vm-host>

Then make Include ~/.containarium/ssh_config the first line of your laptop's ~/.ssh/config. Use the key whose public half was supplied when creating the box: that key is authorized on both the box and its VM jump account. Add --identity ~/.ssh/<your-private-key> to sync if needed; it applies to both SSH connections.

✅ Check:

ssh alice-container hostname   # → alice-container

3. Wire your agent to the box

Your agent reaches the box through agent-box, an MCP server that runs inside it. code install puts it there, along with Claude Code. Run it on the VM, in the same shell as step 1:

containarium code install alice

✅ Check: the install prints agent-box installed at …/.local/bin/agent-box, and the binary is there:

ssh alice 'test -x ~/.local/bin/agent-box && echo agent-box ready'   # → agent-box ready

On the machine where your agent runs, quickstart wires the MCP config automatically. Step 1 already did this for Claude Code on the VM. If your agent runs on your laptop, run this there using the API connection from step 2; it reuses the existing box:

containarium quickstart alice --agent claude
# For Gemini or Codex, use --agent gemini or --agent codex instead.

For Cursor, add the server to ~/.cursor/mcp.json manually:

{
  "mcpServers": {
    "containarium-box": {
      "command": "ssh",
      "args": ["alice", "sh -c 'PATH=\"$HOME/.local/bin:/usr/local/bin:$PATH\" exec agent-box'"]
    }
  }
}

This command sets the box's PATH for non-interactive SSH and supports both user-level and system-wide installs of agent-box.

✅ Check: your agent lists containarium-box as a connected MCP server (in Claude Code, claude mcp list). Asking it to "run uname -a in the box" returns a Linux kernel line from alice-container. It now has shell_exec, read_file, write_file, list_directory, move_file and delete_file inside the box.

That's your first box. Next: run the agent on the box · run pi inside the box · put the box on a public hostname · build a site in one command.

---

After your first box

Run the agent on the box

code install (step 3) already put Claude Code on the box. It installs no credential: sign-in goes through Anthropic's own flow, never through us. Choose one of these, once:

# interactive: sign in inside the box (device code)
ssh alice
claude

# or headless: your own key in the "env" block of ~/.claude/settings.json on the box
#   {"env": {"ANTHROPIC_API_KEY": "<your key>"}}

Then:

containarium code run alice --prompt "add a health endpoint and run the tests"

code run streams output to your terminal as it's produced, but it isn't a pipe. The run is detached on the box and its output is captured to a log; your terminal is a resumable reader over that log. Close your laptop, lose wifi, press Ctrl-C: the run keeps going, and containarium code attach alice picks the stream back up byte-exact, with nothing missing and nothing repeated.

containarium code attach alice   # reconnect; replays what you missed
containarium code status alice   # liveness, and the exit code once it finishes
containarium code stop alice     # reap it; the log stays readable

The difference from the MCP wiring in step 3 is where the agent runs. There, the agent is on your machine and reaches into the box. Here, the agent is on the box, so the work survives your machine sleeping, and the tests run next to the code instead of over a network hop.

Run pi inside the box

If your agent has no MCP client, run it inside the box instead. See docs/integrations/pi.md for the pi walkthrough (installs, keys, code sync, sessions).

Put the box on a public hostname

containarium expose-port alice \
  --container-port 8080 \
  --domain blog.example.com

Caddy on the sentinel terminates TLS for blog.example.com and forwards to alice-container:8080. curl https://blog.example.com then hits whatever alice is serving on port 8080.

Build a site in one command

Run from your laptop against a server you can reach (see step 2, variant B), quickstart does steps 1–3 in one go: it creates the box and wires SSH and your agent (claude, gemini or codex). With --prompt it also exposes --domain and launches your local agent on the build:

containarium quickstart alice --server <your-server> \
  --prompt "a coffee-shop landing page" --domain coffee.example.com

containarium quickstart --help lists every option.

---

The four primitives

Every action in Containarium has a CLI verb (canonical) AND an MCP tool (thin wrapper that delegates to the same Go function). See CLAUDE.md for the convention.

agent-box — in-the-box MCP server

Runs inside every container. Reached over stdio (typically wrapped by SSH on the client side). Exposes Linux-native operations:

Tool	What it does
shell_exec	Run a shell command, capture stdout/stderr/exit, bounded by timeout (default 30s, max 10min) and 256 KiB output cap
read_file	Byte range OR head=N lines OR tail=N lines
write_file	Atomic write with mkdirp (temp + rename)
list_directory	Type/size/mtime, hidden filtering
move_file	Atomic rename with mkdirp on destination
delete_file	Single-file remove (refuses directories so recursive deletes go via shell_exec where blast radius is explicit)

Resources (read-only data the agent fetches via MCP resources/read):

URI	What it returns
containarium://ci-context	JSON metadata about the current CI run (PR number, commit SHA, failing test, etc.) when the box was kept alive by the FootprintAI/containarium-run GitHub Action after a failed CI run. Returns {"available": false} on non-CI boxes so callers never have to special-case errors.
containarium://ci-prompt	Static markdown playbook telling agents how to debug a failing CI run inside this box (what to read first, how to iterate, what not to do). Same body on every box; pair with ci-context for the per-run data.

Optional sandbox: when AGENTBOX_ROOT is set, every file-ops path is resolved against that root with a boundary-aware prefix check. Default unset = no constraint. See internal/agentbox/ for the Go implementation.

mcp-server — platform MCP server

Runs on the host. Exposes the platform's outside-the-box operations as MCP tools — about a hundred of them, one per CLI verb: container lifecycle (create_container, list_containers, start_container, stop_container, delete_container, resize_container), networking (expose_port, sync_ssh_config, passthrough routes), coding-agent runs (code_run, code_attach, code_status, code_stop), secrets, backups, recipes, security scans, tracker connections, and get_metrics / get_system_info. A token's scopes decide which ones a client sees. See cmd/mcp-server/ and internal/mcp/ for the catalog.

containarium CLI

Same surface as the platform MCP, plus deeper administration. The verbs this README uses (the CLI has about ninety — containarium --help lists them all):

containarium quickstart         Create a box, wire SSH + your agent, optionally build a site
containarium create             Create a new container
containarium list               List all containers
containarium wake / sleep       Start a stopped box / stop a running one
containarium delete             Delete a container
containarium code               Run a coding agent ON a box (install / run / attach / logs / runs / status / stop / egress-policy)
containarium expose-port        Expose container:port on a public hostname
containarium ssh-config         Generate self-contained ssh_config (show / sync)
containarium route              Manage proxy routes (low-level)
containarium passthrough-route  Manage TCP/UDP passthrough routes (daemon API)
containarium token              Issue, inspect, refresh and revoke JWT tokens
containarium login / connect    Device-flow sign-in to a server; open a shell on a box with that token
containarium info               Container or system info
containarium version            Print version

Host-side verbs (daemon, sentinel, passthrough, audit, doctor, …) live in containariumd — see Two binaries. Run containarium <verb> --help for full options.

Sentinel — sshpiper + Caddy + PROXY-protocol

The sentinel is a tiny always-on VM (e2-micro on GCP free tier works) that:

• Receives SSH on port 22 (sshpiper routes to the right backend by username).
• Receives HTTPS on 443 (Caddy with TLS-passthrough or PROXY-protocol-aware forwarding to backend Caddy).
• Survives spot-VM termination on the backend with a maintenance page.
• Holds the static IP / DNS A-record so backends can be ephemeral.

See docs/SENTINEL-DESIGN.md for the full design.

---

Architecture

        Agent (Cursor / Claude Code / OpenCode)
            │
            │ JWT (access; tt=access, jti, scopes)
            │ MCP over stdio  ──┐
            │                   │ ┌── refresh ──> POST /v1/tokens/refresh
            v                   ▼ │                  (single-use; old jti revoked)
        ssh user@box  → sshpiper → agent-box (in container)
            │
            │ HTTPS  (mTLS upstream; PROXY-protocol v2)
            v
        Sentinel (e2-micro, always-on)
        ├── sshpiper (port 22)            : routes by username; fail2ban per-user
        ├── Caddy + PROXY-protocol (443)  : routes by hostname / SNI suffix
        └── /wake/ source-IP allowlist    : trusted-proxy only
            │
            v
        +-------------------------------------------------+
        | Backend VM (spot or bare-metal GPU node)        |
        |                                                 |
        |  Incus (LXC) ── containers                      |
        |    ├── alice-container    : SSH + agent-box     |
        |    │   └── /run/secrets/* : tmpfs, 0440 alice   |
        |    └── bob-container      : ZFS-backed storage  |
        |                                                 |
        |  Containarium daemon                            |
        |    ├── JWT auth (iss/aud/jti/scopes)            |
        |    ├── Admin RBAC + container-owner authz       |
        |    ├── Image-digest gate (REQUIRE + VERIFY)     |
        |    ├── Secrets ── Postgres (envelope-encrypted) |
        |    │              │                             |
        |    │              v                             |
        |    │           KMS ── Vault Transit / GCP KMS   |
        |    │           (master key retirable post-cutover)
        |    └── Audit log ── Postgres + SHA-256 hash     |
        |                     chain (verify CLI)          |
        +-------------------------------------------------+

A single sentinel can front multiple backend VMs — a "pool" — and a single deployment can run multiple pools (each with its own backends). See docs/MULTI-POOL.md.

Security control surface (all opt-in via env, default-off for upgrade safety; see docs/security/OPERATOR-SECURITY-RUNBOOK.md):

Env var	Layer	Effect
CONTAINARIUM_REQUIRE_IMAGE_DIGEST=true	API	refuse images without @sha256:<64hex>
CONTAINARIUM_VERIFY_IMAGE_DIGEST=true	API	verify digest against the registry index (pre- + post-pull)
CONTAINARIUM_ALLOWED_IMAGE_REGISTRIES	API	restrict which simplestreams remotes the daemon will pull from
CONTAINARIUM_KMS_BACKEND={none,inproc,vault,gcp}	Secrets	envelope-encrypt DEKs through an external KMS
CONTAINARIUM_REQUIRE_ENVELOPE=true	Secrets	refuse legacy master-key-only rows (Phase E retirement gate)
CONTAINARIUM_POSTGRES_URL_FILE / _PASSWORD_FILE	Secrets	DB creds from disk rather than env
CONTAINARIUM_WAKE_TRUSTED_PROXIES	Sentinel	source-IP allowlist for /wake/
OTEL_BEARER_REQUIRED=true	Telemetry (collector-side; set on the OTel collector, not the daemon)	collector rejects un-bearered OTLP submissions

---

How it's different

vs. SaaS-only sandboxes (e2b, Modal, Replit)

These give you sandboxes for AI agents, but only as hosted SaaS:

• Self-hostability: Containarium runs on your own infrastructure (a $5 VM, your homelab, your enterprise data center). e2b, Modal, and Replit are SaaS-only — your code, your data, and your customers go through their compute.
• License: Apache 2.0, no CLA. Fork it, sell it, run it.
• Surface: full Linux containers with systemd, real network namespaces, GPU passthrough. Not a process-per-call sandbox.
• Transport: MCP-native from day one, not a custom SDK with MCP bolted on.

vs. Docker AI Sandboxes (sbx)

Docker's sbx run claude and Containarium both call themselves "AI sandboxes," but they sit on opposite ends of the same spectrum:

• Locality: sbx runs the sandbox on the developer's laptop (microVM, host-isolation). Containarium runs the sandbox on a VM you host (LXC, multi-tenant, public-internet reachable via the sentinel).
• Persistence: sbx is session-shaped (workspace mount, no documented "give me a box that survives reboot and has a hostname"). Containarium containers persist indefinitely, with ZFS snapshots and 30-day retention.
• Public reach: containarium expose-port alice --domain blog.example.com is one verb. sbx is laptop-local; no public-hostname story.
• Agent surface: sbx is CLI-first (sbx run <agent>). Containarium is MCP-native — two MCP servers (in-the-box agent-box + platform mcp-server) plus the same surface via CLI, SSH, REST/gRPC, and a web UI.
• License: sbx CLI is free; team policy (Docker Admin Console) is a paid subscription. Containarium is Apache 2.0 — including the audit log, RBAC, KMS integrations, and everything else on this page.

If you're stopping an agent from rm -rf-ing the laptop it's running on, sbx is the lighter tool. If you're giving your agent (or your customer's agent) a persistent Linux box on the public internet, Containarium is the shape.

vs. OSS Kubernetes agent runtimes (agent-sandbox, OpenShell)

kubernetes-sigs/agent-sandbox and NVIDIA/OpenShell are the closest open-source peers on Kubernetes:

• SSH-native vs. exec-based: both agent-sandbox and OpenShell reach the sandbox via kubectl exec or a proprietary client, which requires the agent to hold a kube-apiserver token or cluster credentials. Containarium reaches the pod over SSH through sshpiper — the agent has no path to the cluster control plane at all.
• MCP-native: agent-sandbox and OpenShell expose REST APIs or custom SDKs. Containarium's agent-box MCP server runs inside the box, reachable over SSH stdio — any MCP-speaking agent (Claude Code, Cursor, OpenCode) works with zero client library.
• LXC + K8s, one CLI: Containarium runs on either Incus/LXC or Kubernetes behind the same containarium CLI and --runtime flag. You switch backends without changing anything for the agent.
• eBPF egress policy: Containarium can enforce per-tenant egress allowlists at the kernel level via TC_INGRESS eBPF programs once an operator arms them (opt-in; see docs/security/CLAIMS.md). agent-sandbox has NetworkPolicy; OpenShell has eBPF but is NVIDIA-stack-specific. Neither offers a portable, per-tenant eBPF allowlist across LXC and K8s backends.

vs. dev environment platforms (Codespaces, Gitpod, Coder)

Those are persistent IDEs. Containarium is a persistent box — agent-driven, not developer-driven, no IDE assumption, SSH-as-the-API:

• Containarium environments are reached by SSH and MCP. Any IDE works (Vim, JetBrains Remote, VS Code Remote, Cursor's remote dev — your call).
• Cost: no per-hour billing in the OSS path. Self-host costs are just your underlying VM.
• Persistence: containers survive indefinitely; Codespaces auto-delete after inactivity.

vs. application container platforms (Docker, Kubernetes)

LXC is a system container, not an application container. Each container has systemd, a real init, real users, real package managers, real sudo. You can run Docker inside a Containarium container; the reverse isn't really a thing.

If your agent is going to apt install half a Linux distro, edit config files in /etc, run a database, and reboot — LXC is the right shape. If your agent runs a single Python process, Docker or Modal is fine.

It isn't either/or: Containarium can run a box as a pod in a Kubernetes cluster you already operate — same SSH-native agent contract, no kube-apiserver token in the agent's hands. Switch with --runtime=k8s. See the Kubernetes backend section below.

---

What's in the box

Beyond the agent-native primitives, Containarium ships:

Multi-OS

• Ubuntu 24.04 LTS (default)
• Rocky Linux 9 (dev/test)
• RHEL 9 (production)
• Windows Server VMs via QEMU/KVM with RDP — see docs/WINDOWS-VM-SETUP.md

GPU passthrough

For ML/AI agent workflows. Works with NVIDIA RTX 3090, RTX 4090, and similar. PCI-level passthrough so the container sees the GPU directly. Tested on bare-metal GPU nodes connected to the sentinel via tunnel.

Multi-backend

A single sentinel can front:

• GCP spot VMs: cost-effective cloud backends with auto-recovery on preemption.
• Bare-metal GPU nodes: any Linux box you can SSH to; reaches the sentinel via outbound tunnel.
• Windows VMs: live alongside Linux backends.

All containers from all backends appear in a single unified API.

Kubernetes backend (experimental)

Beyond the LXC/Incus backend, Containarium can run a box as a pod in a Kubernetes cluster you already operate — reached over SSH exactly like an LXC box, so an agent can't tell which substrate it landed on. The daemon reconciles a per-tenant namespace + StatefulSet + headless Service

• default-deny NetworkPolicy, and programs the sshpiper gateway (the Pipe CRD) so ssh <tenant>@<gateway> routes to the right pod. What the NetworkPolicy does and does not guarantee is in docs/security/CLAIMS.md.

The pitch isn't "another way to run pods" — it's giving an agent a hardened, SSH-native foothold in your cluster without handing it a kube-apiserver token: the box runs with automountServiceAccountToken: false and satisfies the restricted Pod Security profile. Compare this to kubectl exec-based runtimes where the agent necessarily holds cluster credentials.

Runtime selection (no recompile needed):

# LXC/Incus (default)
containariumd daemon

# Kubernetes — uses in-cluster config or KUBECONFIG
CONTAINARIUM_RUNTIME=k8s containariumd daemon
# or
containariumd daemon --runtime=k8s

Both backends share the same CLI, MCP tools, JWT auth, and REST/gRPC API. GPU passthrough is supported on K8s via nvidia.com/gpu resource limits. Local bring-up takes under 5 minutes with kind — see docs/KIND-QUICKSTART.md.

Design, topology, and BYO-cluster integration: docs/K8S-AGENT-BOX-RUNTIME-DESIGN.md.

Web UI

A basic dashboard at /webui/ for users who'd rather not type CLI: container list, lifecycle controls, metrics, browser-based terminal. Polished UI is intentionally a cloud-product concern — the OSS web UI is functional, not opinionated.

Persistent storage (ZFS)

Containers survive VM restarts and spot termination. ZFS handles compression, snapshots (daily by default, 30-day retention), and checksums.

Sentinel HA

The sentinel itself is e2-micro (free tier). It:

• Detects spot preemption in ~10s, serves a maintenance page.
• Restarts spot VMs automatically (~85s total recovery).
• Holds the static IP, so DNS doesn't change as backends rotate.

Monitoring & observability

VictoriaMetrics + Grafana auto-provisioned. Per-container CPU, memory, disk, network. Alerting via webhooks. SSH audit logs per user (LXC backend; the Kubernetes backend records control-plane API audit but not in-box SSH sessions — see #1189).

Security primitives

• Unprivileged LXC containers: container root ≠ host root.
• Per-user proxy accounts: /usr/sbin/nologin on the sentinel, users can only proxy through to their container.
• fail2ban per-user: an attack on Alice's account doesn't ban Bob.
• Trivy image/dependency scanning across backends. ClamAV malware scanning is a tenant-run capability (the scan runs inside a box, started by its owner, and the verdict is self-reported); it is being rebuilt in #2631 and is not a platform-wide control today.
• AppArmor profiles per container.
• AGENTBOX_ROOT sandbox to constrain agent-box file ops at runtime.

Zero-trust controls (see docs/security/OPERATOR-SECURITY-RUNBOOK.md; several are opt-in, see docs/security/CLAIMS.md for what is on by default):

• JWT with iss / aud / jti / tt / scopes: 32-byte minimum secret enforced at startup; refresh tokens are single-use; jti-based revocation; per-tool MCP scopes propagate to server-side gates.
• Admin RBAC + per-container ownership on the API surface; cluster ops admin-only, container ops owner-only.
• KMS envelope encryption for tenant secrets (Vault Transit or GCP Cloud KMS), with a migration tool and master-key retirement gate.
• tmpfs --delivery=file for secrets that shouldn't be visible in /proc/<pid>/environ.
• Audit log with SHA-256 hash chain + containariumd audit verify (on the host): tamper-evident against an application- or database-level attacker. An attacker who can write the database can rewrite the chain consistently; the chain root can be anchored to a file outside the database (internal/audit/anchor.go), which a database credential does not reach, but a host-root attacker could still edit both.
• Image-registry allowlist + pre-pull simplestreams digest verification + post-pull volatile.base_image defense-in-depth for supply-chain hardening.
• SECURITY.md with a 90-day coordinated-disclosure window; gosec / govulncheck / trivy running in CI.

---

CLI reference (essentials)

Container lifecycle

# Create (Ubuntu 24.04, default)
containarium create alice --ssh-key ~/.ssh/id_ed25519.pub

# Create with options
containarium create ml-dev \
  --ssh-key ~/.ssh/id_ed25519.pub \
  --gpu 0 \
  --stack gpu \
  --memory 16GB \
  --cpu 4

# Lifecycle
containarium list
containarium info
containarium wake alice     # start a stopped box
containarium sleep alice    # stop it
containarium delete alice

Networking

# Expose a container port on a public hostname
containarium expose-port alice \
  --container-port 8080 \
  --domain blog.example.com

# Lower-level route management
containarium route add api.example.com --target 10.0.3.42:3000
containarium route list
containarium route delete api.example.com

# Raw TCP/UDP passthrough via local iptables (on the host, so it's a
# containariumd verb)
containariumd passthrough add --port 50051 \
  --target-ip 10.0.3.150 --target-port 50051

# Raw TCP/UDP passthrough via the daemon API (works against a remote box)
containarium passthrough-route add --port 9443 \
  --target-ip 10.0.3.150 --target-port 50051 --server <host:port>
containarium passthrough-route list --server <host:port>
containarium passthrough-route remove --port 9443 --server <host:port>

SSH config

# Print to stdout (preview)
containarium ssh-config show

# Write to ~/.containarium/ssh_config (one-line `Include` to wire in)
containarium ssh-config sync
containarium ssh-config sync --sentinel sentinel.example.com  # via sentinel
containarium ssh-config sync --identity ~/.ssh/containarium_ed25519

# From a client machine, against a remote server (the bare form above
# only reaches a local Incus socket)
containarium ssh-config sync --http --server <host:port>

Authentication

# Issue an access + refresh pair (CLI-only; never exposed via API).
# Access tokens are short-lived (default 15 min) and authenticate the
# API. Refresh tokens are long-lived and single-use — exchange via
# POST /v1/tokens/refresh for a new pair.
containarium token generate \
  --username admin \
  --roles admin \
  --secret-file /etc/containarium/jwt.secret

# Use the access token
curl -H "Authorization: Bearer <access-token>" http://localhost:8080/v1/containers

# Exchange the refresh token for a new pair (same as POST /v1/tokens/refresh;
# the refresh token is the credential, so no --token)
containarium token refresh --refresh-token "$REFRESH" --server <host:port>

# Inspect a token's claims (jti, scopes, expiry, validation)
containarium token inspect <token> --secret-file /etc/containarium/jwt.secret

# Revoke a leaked token by jti (idempotent; reads from `audit query`
# or `token inspect`)
containarium token revoke <jti> --reason "leak_2026_05_22"
containarium token list-revoked

# Mint a least-privilege token for an agent with only the scopes it
# needs — server-side gates enforce this even if the agent ignores
# the filter.
containarium token generate \
  --username alice-agent \
  --scopes containers:read,containers:write \
  --secret-file /etc/containarium/jwt.secret

See docs/security/OPERATOR-SECURITY-RUNBOOK.md for the full token lifecycle, leak-response playbook, and the agent least-privilege scope catalog.

---

Deployment

Two binaries: containarium vs containariumd

Same convention as docker/dockerd or incus/incusd: containarium is the CLI you run from your laptop (create, list, ssh-config, expose-port, …); containariumd is what runs on the host — the daemon, the sentinel, and every on-host admin verb (daemon, sentinel, service install, sync-accounts, pool join, node, hosting, recover, doctor). Install containarium wherever you want to drive the platform from (your laptop, a jump box); install containariumd on every host that runs the daemon or sentinel itself. A compat symlink (/usr/local/bin/containarium -> containariumd) keeps existing hosts and scripts that still invoke containarium <daemon verb> working during the rollout.

Manual install (recommended for getting started)

curl -fsSL https://containarium.dev/install.sh | sudo bash

See hacks/README.md for what the script does.

Terraform (recommended for production)

cd terraform/gce
cp examples/single-server-spot.tfvars terraform.tfvars
vim terraform.tfvars   # set project_id, admin_ssh_keys, allowed_ssh_sources
terraform init
terraform apply

See terraform/gce/README.md for variables.

System requirements

• Host OS: Ubuntu 24.04 LTS or later (containers can be any supported OS).
• Incus 6.19+ required for Docker-in-LXC support. Ubuntu 24.04's default repos ship 6.0.0 which has an AppArmor bug (CVE-2025-52881); use the Zabbly Incus repository for current builds.
• ZFS kernel module (for disk quotas).
• Kernel modules: overlay, br_netfilter, nf_nat (Docker in containers needs these).

# Quick Incus install via Zabbly
curl -fsSL https://pkgs.zabbly.com/key.asc | \
  sudo gpg --dearmor -o /usr/share/keyrings/zabbly-incus.gpg
echo 'deb [signed-by=/usr/share/keyrings/zabbly-incus.gpg] \
  https://pkgs.zabbly.com/incus/stable noble main' | \
  sudo tee /etc/apt/sources.list.d/zabbly-incus-stable.list
sudo apt update
sudo apt install incus incus-tools incus-client
incus --version  # 6.19 or later

---

API

Containarium exposes:

• REST API at http://localhost:8080 (gRPC-gateway over the gRPC service, JWT auth)
• gRPC at :50051 (mTLS, primarily used by the CLI)
• Two MCP servers: mcp-server (platform) and agent-box (in-the-box)

OpenAPI / Swagger UI at http://localhost:8080/swagger-ui/.

Token-issuance is CLI-only by design; the daemon does not have an "issue token via API" endpoint, because if it did, anyone with API access could mint admin tokens.

---

Hardening notes

SSH key hygiene

• Each user gets their own keypair. Never share keys between users — sharing breaks revocation, audit, and per-user fail2ban.
• The same key can authenticate to both the sentinel proxy account and the container. That's the supported flow: simpler for users, no security loss because the proxy account is nologin and only routes through.
• To rotate: user generates a new key, admin replaces the authorized_keys content in the container.

Agent-box sandbox

If you're running an untrusted agent, set AGENTBOX_ROOT to a project directory:

# In the container
export AGENTBOX_ROOT=/srv/project
agent-box   # all file ops now constrained to /srv/project

shell_exec is intentionally not constrained beyond the LXC container boundary itself — by design, that's the tool's contract. If you need tighter isolation, run agent-box in a more restrictive container (e.g. nested LXC, or chroot the user account further).

Network

• Backend VMs have no public IP by default; they reach out via Cloud NAT and accept inbound only via sshpiper.
• Sentinel allowlist: configure allowed_ssh_sources in Terraform (or firewall rules manually) to lock down who can hit port 22.

---

Comparison FAQ

Why not Docker / Podman? Docker is for application containers. Containarium uses LXC system containers — full Linux OS per container, real systemd, native SSH, Docker-in-LXC works, persistent filesystem. If your agent will apt install and reboot, you want LXC.

Why not Kubernetes? K8s orchestrates application containers across nodes — that's the infrastructure layer. Containarium is the agent runtime that runs on top of Kubernetes (or LXC), giving each agent a persistent, SSH-native box without handing it cluster credentials. If you're already on K8s, run containariumd daemon --runtime=k8s and use your existing cluster as the backend. See the Kubernetes backend section and docs/KIND-QUICKSTART.md.

Why not Vagrant? Vagrant orchestrates VMs on a developer's local machine. Containarium hosts environments on shared remote infrastructure for many agents.

Why not Dev Containers / VS Code Remote Containers? Dev Containers are project-scoped, IDE-coupled, single-developer. Containarium gives many users (or many of one user's agents) their own persistent boxes on shared infrastructure, IDE-agnostic.

Why not Codespaces / Gitpod? Browser-IDE-as-a-Service, per-hour billed, vendor-locked. Containarium is self-hosted, persistent, SSH/MCP-based, no per-hour billing in OSS.

Why not e2b / Modal / Daytona? Closest peers — sandboxes for AI agents. They're SaaS-only and typically optimize for short-lived, process-per-call execution. Containarium is self-hostable, MCP-native, and gives you full persistent Linux boxes. Pick e2b if you want hosted-only and ephemeral; pick Containarium if you want self-hosted, persistent, and your data on your infra.

Why LXC at all?

• Each container runs a full Linux OS with systemd.
• SSH access is first-class.
• Docker-in-LXC works (vs. fragile Docker-in-Docker).
• Real persistent filesystem, real users, real sudo.
• "Feels like a VM" for the agent — same surface area as a managed cloud VM, fraction of the resource cost.

---

Use cases

• AI-agent sandboxes (the lead): Cursor, Claude Code, Cline, OpenCode, custom agents — all reach the same MCP surface.
• Shared developer environments: many developers, one host, SSH jump server with per-user accounts and containers (isolation caveats: docs/security/CLAIMS.md).
• ML / GPU experimentation: GPU passthrough into LXC.
• Education, bootcamps, workshops: per-student isolated Linux with no per-student VM.
• CI / build infrastructure: long-lived build hosts that keep caches warm across runs.
• Demo / testing infrastructure: spin up a real Linux env, test, tear down.

---

In the wild

Where Containarium has been demonstrated live:

• 2026-08 — COSCUP, Taipei. An SSH-native agent runtime, built because our hardware was sitting idle — where utilisation goes on AI and VM fleets, why one VM per developer stopped paying for itself, and how LXC + sshpiper + persistent disk became Containarium. Slides.
• 2026-06-04 — AI Agent Night, Taipei. When the grand-prize giveaway (a custom vibe-keyboard) hit a snag — the event had no way to run the lucky draw — we fired up Containarium and vibe-coded a lucky-draw picker on the spot to save the giveaway. Still live: lucky-draw.demo.containarium.dev.

Demoed Containarium somewhere? Open a PR and add it here.

---

Status

• Production-deployed on GCP (multi-region) and bare-metal GPU nodes.
• APIs are stable (protobuf-defined with gRPC-gateway).
• Apache 2.0, no CLA, accepting community PRs.
• Active maintenance: see commit history on main and recent releases.

---

Roadmap

• Shipped (2026-06): Kubernetes backend — run a box as a pod in a cluster you operate, reached over SSH like an LXC box, with the sshpiper gateway, default-deny NetworkPolicy, CSI storage, and nvidia.com/gpu passthrough. Selected at runtime via --runtime=k8s (no recompile). End-to-end validated on kind; 5-minute local bring-up in docs/KIND-QUICKSTART.md. See also docs/K8S-AGENT-BOX-RUNTIME-DESIGN.md.
• Shipped: agent-box MCP, ssh-config CLI, expose-port CLI, and the demo recording — the Quick start above is built on them.
• Q3 2026: agent-box tier-2 (MCP Roots, background process management), demo-driven docs and examples.
• Q4 2026: OSS v1.0 cut — stable API surface, contribution guide.

If you want to drive an item, open an issue or PR — community work is welcome and we triage weekly.

---

Contributing

• Read CLAUDE.md for the CLI-first principle (every new platform action lands as containarium <verb> first; MCP wraps it).
• Check existing issues and PRs.
• Add tests for new features.
• Update docs if user-visible behavior changes.

No CLA. Apache 2.0 means you can use, modify, and redistribute. We welcome PRs that align with the project's positioning and reject those that don't (e.g. "let me add multi-tenancy to the OSS daemon" goes into the cloud repo discussion, not here).

---

License

Apache License 2.0 — see LICENSE.

---

Acknowledgments

• Incus — modern LXC manager.
• sshpiper — SSH reverse proxy.
• mcp-go — Go MCP server library.
• Caddy — TLS / reverse proxy with PROXY-protocol support.
• Cobra — CLI framework.
• Terraform — infrastructure as code.

---

Support

• Project site: containarium.dev — overview, hosted cloud, GitHub Action for CI, PR previews.
• Documentation: docs/ directory.
• Issues: GitHub Issues.
• Demo video: 55s walkthrough on YouTube.
• Live demo app: helloworld.demo.containarium.dev.
• Organization: FootprintAI.