Orchestrate multi-agent AI workflows for full-stack development

Maestro-Flow is an intent-driven, multi-agent workflow orchestration framework: instead of driving one agent through one task, its Ralph engine reads project state, classifies a natural-language intent into one of 40+ command chains, and coordinates multiple AI CLIs through the whole cycle from brainstorming to deployment - deciding at each checkpoint whether to continue, roll back, or insert a fix loop. A knowledge graph automatically persists patterns, pitfalls, and decisions discovered during execution (as Spec and Knowhow) and injects them into later agents' prompts via a hook system, so a project's own history makes later runs smarter.

Use it for multi-step development work that benefits from an adaptive pipeline rather than a fixed script: implementing a feature end-to-end (its own example is OAuth2 with refresh-token support), long-running debugging or deep refactors via the Odyssey long-cycle autonomous loops (which run for hours and adapt strategy at each checkpoint until acceptance criteria are met), or coordinating multiple AI backends - Claude, Codex, Gemini, Qwen, OpenCode, and Grok - in one workflow through four orchestration modes (Delegate for async hand-off, Team for role collaboration, Wave for dependency-based parallelism, Swarm for ant-colony-style exploration). For a quick, low-stakes fix (a typo, a small correction) its lightweight /maestro-companion entry point runs a minimal single-turn lifecycle instead of the full pipeline. It requires Node.js 22.19+ and at least one host CLI (Claude Code by default, and/or Grok Build); Grok support specifically needs a separate install script run from the repo root, and a version mismatch against the official package fails outright rather than silently downgrading.

Input is a natural-language intent (/maestro-ralph "implement OAuth2 authentication with refresh token"), which Ralph classifies and expands into a command chain - typically analyze to plan to execute to verify to review to test, with debug/fix/retry loops inserted automatically on failure. Three quality modes trade thoroughness for speed: full adds business-test and test-gen for production/security-critical work, standard is the default balance, and quick drops straight to a CLI review for prototypes and hotfixes. Output is completed, verified code changes plus persisted knowledge (Spec/Knowhow entries) that future runs draw on; a session can be paused at a decision point and resumed (/maestro-ralph -c) or run fully unattended (-y).

npm install -g maestro-flow@0.5.82
maestro install

It ships an MCP server (stdio) with 9 endpoint tools, and a governed exact-search mode (maestro search "..." --exact) built on @vscode/ripgrep for locating known strings/paths/identifiers without going through the default BM25/embedding-ranked search fusion. The project is roughly 333 TypeScript source files (~80k lines) built on Commander.js, the MCP SDK, better-sqlite3, web-tree-sitter, and a React 19/Vite dashboard, and credits GET SHIT DONE for its spec-driven methodology and its own predecessor project, Claude-Code-Workflow.

Developers running multi-step or multi-day development work across more than one AI CLI who want an adaptive pipeline - one that decides when to retry, roll back, or escalate - rather than a fixed script, and who want the project's own accumulated knowledge fed back into every subsequent agent run. Maestro Flow is released under the MIT license.