pve-agents — self-hosted sandboxes for coding agents

Give it a repository, a ref and a purpose. It builds a disposable LXC on your Proxmox host, starts a Claude Code agent inside it, and streams the whole thing to your browser.

repondom91/pve-agents

refmain

purpose

01clonegolden template → new LXC
02bootcontainer up, address assigned
03checkoutrepository at your ref
04agentClaude Code, handed the purpose

Run as many as your host will hold.

Each workspace is its own container, so nothing an agent does touches your workstation. Every one starts from the same clean template, they run in parallel rather than in turn, and the hardware is already yours.

The fleet: one workspace running, forty-seven destroyed, and the form that requests the next one

What is running, what has been destroyed, and the form that starts the next one.

The loop

Watch it work, answer it, keep or throw away the result.

Watch

The transcript streams

Tool calls and all, over SSE, replayed from the beginning for whoever opens the page late.

Answer

Its questions surface in the browser

Permission prompts are answered where you already are, not in a terminal somebody has to be attached to.

Decide

Read the diff, then push or discard

Per-file diffs in the rail. Push to a branch of the workspace's own, or throw the lot away.