Open-source · Apache 2.0 · self-hostable

Run AI-generated code — safely, in an isolated box.

Code an agent writes is code you didn't review. Run it in a disposable Linux box — isolated from your machine and every other tenant — instead of on your laptop or a shared host. Self-host it so the code never leaves your network.

~ / containarium
# spin up a throwaway box, run the generated code inside it
$ containarium create run-box
 isolated box run-box ready
$ containarium ssh run-box -- 'python /tmp/agent_output.py'
…runs in isolation, not on your machine…
$ containarium delete run-box   # gone, no trace

Why isolate it

Generated code is untrusted code.

An agent can write a script that deletes the wrong directory, exfiltrates a secret, or pulls a malicious dependency. On your machine that's a real incident. In a Containarium box it's contained.

Isolated & disposable

One LXC box per agent or per task, sandboxed from the host and other tenants. Delete it when the run is done — no residue on your machine.

Network deny-by-default

Optional eBPF per-tenant policy blocks egress except a CIDR/domain allowlist, so generated code can't quietly call home or reach the rest of your network.

Stays in your perimeter

Self-host on your own VM and the code, data, and artifacts never leave your network — no shipping untrusted code to a third-party cloud.

How it works

Box up, run, throw away.

  1. STEP 1
    Provision a box

    One command spins up an isolated Linux box with its own filesystem and network.

  2. STEP 2
    Run the code

    The agent executes generated code inside the box over MCP or SSH — and can inspect logs and iterate in place.

  3. STEP 3
    Keep or discard

    Keep the box as a persistent workspace, or delete it — the blast radius dies with it.

FAQ

Common questions

How do I run AI-generated code safely?

Run it in an isolated, disposable box instead of on your machine. Containarium gives each agent its own LXC container, sandboxed from the host and other tenants, with optional eBPF deny-by-default network controls. Execute there, then tear it down.

Where should AI agents run the code they generate?

Not on a laptop or shared host — one isolated box per agent or task, so a bad install or destructive command is contained. Containarium provisions that box in one command and exposes it over MCP.

Can I run it without sending code to a third-party cloud?

Yes — the core is open source (Apache 2.0) and runs on your own VM, so code and data stay in your network. A hosted option exists if you'd rather not run the infrastructure.

How is this different from a serverless code-execution sandbox?

Serverless sandboxes reset between calls. A Containarium box is a persistent full Linux environment with root and a routable hostname — install tools once, keep caches warm, host what it builds — while still isolated and disposable when you want. See the sandbox comparison.

Run it where it can't hurt you.

Start free on the hosted cloud, or self-host the open source on your VM.