Remote execution for developers allergic to CI waits

What it does

Crabbox is a remote execution control plane that leases a machine, syncs your dirty working tree over SSH/rsync, runs the command, streams output back, and releases the box. It targets the gap between local terminal convenience and CI rigidity: you keep editing locally, but the heavy lifting happens on cloud capacity, an existing SSH host, or a delegated sandbox. A Go CLI handles the data plane directly to the runner, while an optional coordinator manages provider credentials, lease state, spend caps, and cleanup.

The interesting bit

The project explicitly refuses to be CI, a security sandbox, or a secrets scrubber—its trust model assumes cooperative team members, not adversarial tenants. That honesty is refreshing; instead of overpromising isolation, it focuses on evidence-producing runs with captured output, logs, and artifacts for human reviewers or AI agents. The coordinator itself is deployment-agnostic, running either on Cloudflare Workers with Durable Objects or as a Node.js service backed by PostgreSQL, though the Node runtime is newly shipped and still needs production validation.

Key highlights

• Direct SSH/rsync data plane from the CLI to the runner; the coordinator only brokers leases and credentials, not traffic.
• Supports provider lifecycle delegation for AWS, Azure, GCP, Hetzner, and Daytona; everything else runs direct or delegated without coordinator transfer.
• Optional coordinator runtimes: established Cloudflare Workers stack, or a newer Node.js/PostgreSQL path that requires validation before cutover.
• Integrates with coding agents via a generic Skill, Zed Remote Projects, and the Herdr action palette.
• Records run evidence and artifacts, but does not automatically scrub secrets from captured output—review before sharing.

Caveats

• The Node.js/PostgreSQL coordinator is newly shipped and not yet the established deployment; the README warns to validate it against a production proof checklist before cutover.
• Not a replacement for CI and explicitly not designed for mutually untrusted tenants; repository configuration is treated as trusted automation.
• Some integrations, like the Zed registry submission and Herdr marketplace indexing, are tracked but not yet completed.

Verdict

Worth a look if your local builds are too slow or you need disposable, matching environments for tests and reviews. Skip it if you need hardened multi-tenant isolation or a fully managed CI replacement.