// the take
A JSON-driven orchestration layer for Claude Code that coordinates multiple AI CLIs (Gemini, Codex, Qwen) through skills, slash commands, a team-based multi-agent architecture, and a queue scheduler with its own dashboard. It's aimed at developers who want to run several AI coding assistants together on complex multi-step work rather than just scripting one.
The orchestration layer is more substantial than typical prompt-template wrappers: a message-bus team architecture, a background queue scheduler with its own API, and session lifecycle management (start/resume/sync/complete) for long-running work. It's not locked to one vendor - CLI selection across Gemini/Codex/Qwen is semantic, triggered from natural language in the prompt rather than hardcoded flags. It ships an actual frontend (multi-terminal dashboard, React Flow-based visual workflow editor) instead of being CLI-only, which is unusual for this category.
The repo is archived - the README's first line states development has moved entirely to a new project (maestro-flow) and this one gets no further updates, so anyone adopting it is building on something already deprecated by its own author. Running multiple autonomous coding CLIs in parallel against a real codebase with file-system access is a lot of blast radius, and the README says nothing about sandboxing, approval gates, or recovery when one of the orchestrated agents does something wrong mid-task. Getting basic functionality working requires manually enabling experimental feature flags on the backing CLIs (e.g. Codex's multi_agent_v2, enable_fanout), which means it's fragile to upstream changes in tools it doesn't control. The CodexLens local search component is explicitly flagged as unstable, and the README is mostly diagrams and badges with little on what failure modes look like in practice.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →