Worktrees that live in the cloud, sleep like serverless, preview like Vercel. Your coding agent enters a worktree — and it's already running on a remote workspace.
Parallel coding agents burn RAM and CPU — installs, dev servers, language servers, test suites — not disk. So the worktrees move to a server, and your machine goes back to being a keyboard.
A Claude Code hook fires when the agent starts a task. The remote workspace exists before you've finished the prompt.
Installs, tests, dev servers — all remote, in an isolated container per branch. Session lives in tmux; your terminal just attaches.
Idle workspaces power off automatically. Disk kept. Preview URL kept. Cost drops to storage.
Reopen the task — or hit the preview URL — and the same workspace resumes where it left off. Not a restart. A resume.
Keep fifty branches alive. Pay for three.
| State | What exists | Cost |
|---|---|---|
| cold | branch + worktree on disk | ~ disk |
| hot | agent + installs + dev server running | metered |
| sleep | container stopped · disk kept · URL kept | ~ $0 |
| gone | wiped after merge + TTL | none |
No new workflow. Claude Code hooks (Pi and OpenCode next) make remote the default — you keep thinking in branches, not machines.
Sleep doesn't kill your agent's session. Wake the workspace and the conversation, the dev server, the shell — all still there.
Worktrees share one repo. Spinning up a parallel task costs megabytes and seconds, not a full VM and a clone.
Every workspace gets a stable URL that survives sleep and wakes on request. Any port. Share it in the PR.
Max hot slots, idle timers, TTLs, budget caps. The bill is a setting, not an accident.
Open source. One binary, one box, your keys, your repos. The hosted version is for people who'd rather not.
Sell you infinite always-on machines? No. Infinite workspaces, finite hot compute — that's the whole trick.
Dogfooding now. Early access when it's boring and reliable.
No spam. One email when early access opens.