Multi-Agent

It runs coding sessions rather than model-call graphs, your code and limits pick the next step instead of a manager model, and scope covers many agents with a shared mailbox rather than one session or hand-managed terminals.

State is stored in SQLite before anything runs, so nothing is resent or marked done by guesswork. Unknown outcomes wait for explicit settlement and a result counts as done only after a person or registered check accepts it.

Yes. Each task gets its own mode, network setting, and readable or writable paths with sign-in untouched, and your application can review each command before it runs.

It really runs keyless. The TypeScript quickstart uses a fake runtime, so it needs no account, no npm install, and makes no model calls; it prints two completed tasks and shows the second reusing the first agent session. The Python version needs 3.11 or later and starts the same Node host. Real Claude Code runs add claude auth login plus the repository dependencies for the Claude Agent SDK. Expect an experimental warning at startup, since it uses Node built-in SQLite. Real-model runs, sandbox enforcement, and Windows are still unverified.

Tasks, messages, approvals, and token usage are written to SQLite before anything runs, so after a crash work whose outcome is unknown waits for you rather than being retried.

Only after a person or a check you registered accepts it; your code and the limits you configure decide what runs next rather than a manager model.

Alpha, tested offline on macOS and Linux with Node 22 and 24 and Python 3.11 and 3.14; runs with real models, OS sandbox enforcement, Windows, and speed or cost comparisons are not verified yet.