Crabbox 0.4.0 launches ephemeral agent machines on Spot instances

• steipete's 0.4.0 launch post frames Crabbox 0.4.0 as "machines for agents on the fly," with AWS Spot, Hetzner, and Blacksmith-backed boxes that can be spun up for short-lived runs.
• The official Crabbox docs describe the same core loop as local edits plus remote execution: sync the dirty checkout, run remotely, stream logs back, then release or keep the lease warm.
• steipete's launchd reproduction demo shows the new pitch in practice, a Codex run validating a macOS-only launchd bug that was hard to reproduce on a non-fresh install.
• Crabbox's current architecture is a Go CLI plus a Cloudflare Worker coordinator and leased machines, according to the architecture docs, with the broker enforcing lease state, credentials, and usage guardrails.

You can read the main docs, skim the architecture page, and browse the GitHub repo. The interesting wrinkle in the 0.4.0 post is the jump from remote Linux test boxes to ephemeral machines for agent runs, while the macOS validation example and the repo's mention of macOS, Linux, and Windows archives make the cross-platform debugging angle much more concrete than the launch copy.

Crabbox started public life as remote Linux test boxes for overloaded laptops. steipete's 0.1.0 post pitched warm boxes, dirty checkout sync, and idle auto-free. By 0.4.0, the framing has shifted to disposable machines for agent workflows.

The official repo README still describes the product in plain terms: lease a machine, sync a working tree, run commands over SSH, and release it. What changed in the new launch post is the emphasis on fast ephemeral capacity for Codex-style runs instead of just offloading local tests.

The useful implementation detail is the broker. The docs homepage says the CLI talks to a Cloudflare Worker that holds provider credentials and lease state, while the coordinator docs list cost guardrails, usage stats, lease heartbeats, and provider resource management as broker responsibilities.

That gives Crabbox a different shape than a plain SSH helper:

• Local editor and git workflow stay on your machine, per the docs.
• The broker owns credentials and serializes fleet state, per the coordinator docs.
• Compute can come from Hetzner, AWS Spot, or Blacksmith Testboxes, per the README and steipete's 0.4.0 post.
• Warm boxes can be kept around and reused with --id, per the docs.

The strongest evidence for 0.4.0 is not the feature list, it is the bug repro. In steipete's demo, Codex uses a Crabbox-backed macOS target to rerun launchd tests, confirm a fresh Darwin environment, and pass both unit and real integration tests.

That lines up with the broader system design in the architecture docs, which describe workers as Hetzner or SSH-accessible machines, not just one fixed cloud image. It also explains why Steinberger's post talks about recreating conditions on macOS, Linux, and Windows: the value is less "remote CI," more disposable environment recreation for agents and maintainers.