Give it an outcome. It detects your project, assembles the right crew, and shows its reasoning before it acts.
Every routing decision is inspectable. Which role resolved to which provider, on which model, and why. The crew manifest shows the plan before any work begins.
Composes the tools you already have and falls back to its own built-ins. Supported lanes include Claude Code, Codex, Kimi, and Cowork, with each harness's limits documented.
Not just code. Product, business, content, and operations pipelines are first-class. PRDs, roadmaps, launch plans, runbooks, and more.
A deterministic Node CLI owns routing, scoring, and validation. No LLM calls. Reproducible by construction. The model is reserved for judgment.
Plan and Go handle one outcome; Plan-backlog and Go-backlog handle batches. Diagnose fixes one bug. Audit sweeps the codebase. Design handles human-facing work. Runner picks one scheduled item. Capture writes approved backlog items. Init prepares repository state.
The CLI makes no model calls. Native Claude Code, Codex, Kimi, and Cowork lanes use their active harness. Private/local ChatGPT Work is the exception: it uses Secure MCP Tunnel, a separately billed OpenAI Platform key, and proof-gated native invocation.
