Persistent Ubuntu VMs for AI agents: SSH/SCP, Docker inside, dedicated IPv6/IPv4 per machine, a 60fps virtual desktop, and disk-level forking from snapshots. CLI and HTTP API with Python/TypeScript SDKs, `--json` on every command for stable JSONL output, `boat prompt` to run Claude Code or Codex inside. Docker, VS Code, Chrome, GitHub CLI, Rust, Node, Bun pre-installed. Pricing is the actual pitch: billed per second, $20/month minimum, which buys 2,000,000 VM-seconds of 4 vCPU / 8 GB time — about 555 hours. Small (2/4) burns at 0.5x, large (8/16) at 2x, so a mixed fleet is just additive. `boat stop` snapshots and pauses billing, which matters more than the headline rate. Self-serve plans scale to 2,000 concurrent sandboxes; default TTL is 1 hour, overrideable. EU regions only for now — Germany, Finland, France — so check latency before you commit a fleet. Their own benchmark shows a Node build/test loop, 1,000 runs, at 36s versus E2B's 89s on identical 4/8 machines. The docs publish an `llms.txt` index with every page as markdown, which is a smart concession: the marketing page admits you shouldn't read it. The comparison table with Cloudflare footnoted for disk resets on sleep is the honest kind of pricing-page knife-fighting. Y Combinator-backed, built by ASCII. Worth noting nobody's benchmark of themselves ever lost. Open question: how does the 1-hour default TTL interact with the "long-running agents, 24/7 sessions" use case the pricing calculator is built around?