Lullabeast
An open-source, local-capable pipeline where a team of planner, executor and reviewer agents builds a working MVP from a plain-English description, phase by phase.
Lullabeast takes a plain-English project description and has agents implement it phase by phase against a real git repository, with gate scripts checking every step and an escalation path back to the user when the agents get stuck. It runs on and requires OpenClaw: Lullabeast provides the pipeline and dashboard, while the agents execute inside OpenClaw's runtime, and a bundled Docker image ships a pinned OpenClaw so installation needs only Docker and an API key.
It is built around known failure modes of cheaper open-weight models: every file deletion must be declared up front or the gate flags it and reruns the phase, context is cleared between tasks to limit spec drift, and a phase cannot close until the gate confirms required tests exist and pass. Each agent gets three attempts with timeouts before a stuck run is flagged to the user, and several projects can be queued and worked in dependency order.
The author describes it as an early release, reliable for small single-purpose web apps, meant to run single-user on a trusted machine behind a locally generated token.