Orchestrate multi-agent AI workflows for full-stack development
Intent-driven multi-agent workflow CLI - describe a goal, and Maestro plans, schedules across Claude, Codex, and Gemini, executes, and verifies it.
Why it matters
Coordinate multiple AI coding agents (Claude, Codex, Gemini) through adaptive lifecycle pipelines that automatically route tasks, execute in parallel, verify quality, and self-heal through debug loops-taking projects from brainstorm to production-ready code.
Outcomes
What it gets done
Route development intent to optimal command chains across brainstorm, blueprint, analyze, plan, execute, and verify phases
Run parallel agent execution with wave-based coordination for independent tasks and dependency management
Insert adaptive debug-fix-retry loops at decision checkpoints based on verification outcomes
Track project state through real-time dashboard with Kanban boards, Gantt timelines, and issue closed-loop management
Install
Add it to your toolbox
Run in your project directory:
curl -fsSL https://spark.entire.vc/get/catlog22-maestro-flow | bash Overview
Maestro Flow
A TypeScript CLI framework that orchestrates multi-agent development workflows from a single stated intent: an adaptive Ralph engine classifies the request into one of 40+ command chains, coordinates Claude, Codex, Gemini, Qwen, and OpenCode agents across four orchestration modes, and persists what each run learns into a knowledge graph that improves future runs. Use it to turn a described goal into a planned, executed, and verified multi-step workflow, or to run long autonomous loops such as debugging, refactoring, or security review via Odyssey, instead of hand-sequencing agent commands and pipeline YAML.
What it does
Maestro-Flow is an intent-driven, multi-agent workflow orchestration framework: state a development goal in natural language, and its Ralph v2 engine reads the project's current state, classifies the intent into one of 40+ command chains, and dynamically decides at each key checkpoint whether to continue, roll back, or insert a fix loop, no YAML pipeline configuration required. It coordinates across backends, mixing Claude, Codex, Gemini, Qwen, and OpenCode within a single workflow through four orchestration modes combined as needed: Delegate (asynchronous handoff), Team (role-based collaboration), Wave (dependency-ordered parallel execution), and Swarm (ant-colony-style exploration). Patterns, pitfalls, and decisions discovered during execution are automatically persisted as Specs and Knowhow, and a hook system injects that accumulated knowledge into subsequent agents' prompts, so the project gets more capable the more it's used.
When to use - and when NOT to
Use the main Ralph entry point, /maestro-ralph followed by an intent, for standard feature work: it builds a chain such as analyze, plan, execute, verify, review, test automatically, inserts a debug, fix, and retry loop on failure, and prepends brainstorm and blueprint steps for a new project. Use the Odyssey command family for large-scale debugging, deep refactoring, or UI optimization that benefits from hour-scale autonomous iteration rather than a single pass - six modes cover debug, planex (plan-execute), improve, review, security (a read-only layered audit covering OWASP, dependencies, secrets, and STRIDE), and ui. Lighter entries exist for smaller asks: /maestro-next routes an intent to the right level of effort automatically, and /maestro-companion handles a minimal, single-step fix like a typo without spinning up a full chain. It requires Node.js 18+ and the Claude Code CLI at minimum, with Codex CLI or agy CLI needed only for multi-backend workflows.
Inputs and outputs
Input is a natural-language intent passed to one of the slash commands: /maestro-ralph, /maestro, /maestro-odyssey, /maestro-next, or /maestro-companion. Output is a fully executed and verified workflow run, tracked through session state that can be paused and resumed, using maestro session status, /maestro-ralph -c to resume from a decision point, or -y to run fully unattended. Three quality modes control how deep the verification pipeline goes: full, verify, business-test, review, test-gen, test, for production or security-critical work, standard, verify, review, test, as the default, and quick, verify, CLI-review, for prototypes and hotfixes.
Integrations
Built with Commander.js and the MCP SDK, backed by better-sqlite3 for its knowledge graph and web-tree-sitter for code analysis, with a React 19, Zustand, Tailwind CSS 4, Hono, and Vite dashboard. The project itself spans roughly 80,000 lines across 333 TypeScript source files, 64 slash commands, 45 skill packs, 23 agent definitions, and 92 templates. It ships an MCP server with 9 documented tool endpoints, and it credits GET SHIT DONE for its spec-driven development approach and Claude-Code-Workflow as its predecessor project for multi-CLI orchestration.
Who it's for
Development teams and solo engineers who want a single intent-driven entry point to plan, execute, and verify multi-step work across several AI coding agents, with accumulated project knowledge automatically feeding back into future runs. The project is MIT licensed.
Discussion
Questions & comments · 0
Sign In Sign in to leave a comment.