Harness
systemctl for your agents — supervise, attach to, and hop between long-running agent CLIs, REPLs and watchers from one Go binary.
🖥️
Keyboard-driven dashboard
Open the cockpit TUI and hop between every harness with live terminal previews — no args, just harness.
♻️
Real supervision
Restart policies, escalating crash-loop backoff and state persistence — harnesses survive client disconnects and daemon restarts.
🤖
Agent adapters
First-class Claude Code, Crush and Codex support: prompt one-shots, model selection, turn budgets, trajectory harvesting.
⏰
Scheduled one-shots
Give a prompt harness a cron schedule and the daemon fires it on cadence — overlapping runs skip, never stack.
📦
Project compose
A repo-root harness.toml gives every checkout its own fleet: harness up / down / ps with project-scoped names.
🔒
Remote over SSH
The same dashboard over the network via Wish — public-key auth only, per-key read-only scoping, no password path.
Under the hood
This project has 18 architecture decisions and 11 specifications recording why every behavior above is the way it is.